A power industry training system based on an NLP large model

Abstract

The application relates to the field of industry training and discloses a power industry training system based on an NLP large model, which is used for providing a personalized training solution for the power industry. The power industry training system comprises the following steps: processing thermal power multi-professional and full production stage standard files by using an NLP large model, establishing a technical supervision standard data source, collecting trainee information to form basic information, generating a customized question bank according to the basic information and matching standard specifications, combining examination and interaction data to generate an ability portrait and a label, and then formulating a personalized training scheme. The application can effectively solve the problems of insufficient personalization, incomplete evaluation and the like of traditional power training, and improve the training effect and safety.

Description

Technical Field

[0001] This invention relates to the field of industry training, and in particular to a power industry training system based on a large NLP model. Background Technology

[0002] In the power industry, especially in the thermal power sector, the technical supervision standards system is vast and complex, covering multiple disciplines and the entire production process. It involves various professional fields such as boilers, steam turbines, electrical systems, and chemicals, and encompasses every stage of the entire lifecycle, from equipment selection and installation to commissioning, operation, maintenance, and testing. With the continuous development of power industry technology and increasingly stringent regulatory requirements, accurately understanding and adhering to these technical standards is crucial for ensuring safe power production, improving operational efficiency, and reducing equipment failure rates.

[0003] Meanwhile, the power industry has a complex personnel structure, with significant differences in the level of technical standards mastery and needs among individuals with different educational backgrounds, majors, professional titles, and work experience. Traditional training methods often adopt a "one-size-fits-all" approach, failing to provide personalized training content based on the actual situation of each individual. This results in inconsistent training effectiveness, with some personnel lacking a thorough understanding and firm grasp of key standards and regulations, which can easily lead to problems such as violations of regulations or insufficient ability to handle emergencies in actual work.

[0004] With the rapid development of information technology, big data, artificial intelligence, and other technologies are widely used in various fields. In the field of power training, how to utilize these advanced technologies to efficiently process and analyze massive amounts of standard documents, accurately assess personnel competence levels, and provide more targeted and effective training programs has become a significant challenge facing the current power industry training work.

[0005] Therefore, we propose a power industry training system based on a large NLP model to address the above problems. Summary of the Invention

[0006] This invention provides a power industry training system based on NLP large models, which is used to provide personalized training solutions for the power industry.

[0007] The first aspect of this invention provides a power industry training system based on an NLP large-scale model. The system includes: a data module for processing standard documents across multiple specialties and the entire production stage of thermal power using an NLP large-scale model to establish a vector database and obtain a technical supervision standard data source; a personnel module for collecting data on trainees' education, major, professional title, and work experience to form basic personnel information; a standards module for matching the basic personnel information with the standards and standards that trainees should master in their specific power generation technology fields from the technical supervision standard data source, generating a customized question bank; a dashboard module for obtaining trainees' test results on the customized question bank and, combined with their interaction data in daily question-and-answer scenarios, analyzing their technical field breadth, activity time, and ability level to generate a capability profile dashboard and capability tags; and a solution module for identifying trainees' capability weaknesses based on the capability tags and generating personalized training solutions including targeted learning content.

[0008] Optionally, in the first implementation of the first aspect of the present invention, the method includes: using at least two different parsing algorithms to perform parallel recognition on the original standard documents, outputting multiple sets of parallel recognition texts respectively; comparing the multiple sets of parallel recognition texts, identifying the parts with consistent content as initial screening texts, and marking the parts with inconsistent content as items to be manually reviewed; merging them after review to form accurate text data; parsing the standard document numbers and reference clauses in the accurate text data to establish a standard reference relationship graph describing the mutual calling relationships between standard documents; traversing the standard reference relationship graph, checking whether the standard document numbers referenced in the accurate text data exist in the graph nodes, removing text paragraphs that reference invalid nodes or non-existent nodes, and generating logically cleaned standard data; segmenting and vector encoding the standard data, generating standard feature vectors and storing them in a database to form a technical supervision standard data source.

[0009] Optionally, in the second implementation of the first aspect of the present invention, the method includes: pre-constructing a classification tree including all technical supervision process nodes in the power industry, and associating each classification node with keywords to form a job-professional mapping table; inputting the collected professional and work experience data of trainees into the job-professional mapping table for keyword matching, filtering out the technical supervision business branches corresponding to the trainees, and generating a target supervision area list; pre-setting a weight matrix based on professional title level and years of work experience, assigning corresponding initial difficulty coefficients to different combinations of professional titles and education, and forming a job level quantification rule table; substituting the education and professional title data of trainees into the job level quantification rule table to find the corresponding initial difficulty coefficients, and binding them with the target supervision area list to form basic personnel information.

[0010] Optionally, in a third implementation of the first aspect of the present invention, the method includes: dividing the acquired data into test result data reflecting passive verification behavior and daily question-and-answer query logs reflecting active exploration behavior; parsing the daily question-and-answer query logs to extract technical keywords actively retrieved by trainees, forming an exploration intent set; projecting the knowledge points of incorrect questions in the test result data and the keywords in the exploration intent set onto the classification tree structure of the technical supervision standard data source; calculating the ratio of the number of leaf nodes actually triggered by trainees to the total number of nodes in the target supervision domain, generating a technical field breadth index; applying a sliding time window algorithm to the daily question-and-answer query logs to calculate the data within a unit time window. The system calculates the effective interaction frequency and marks the timestamp of the last effective interaction to generate an active time trend chart. It reads the test results data of the trainees, calls the difficulty coefficients in the job level quantification rule table, and performs a difficulty-weighted summation of the scores for correctly answered questions to generate a quantified ability level value. It establishes a three-dimensional coordinate system including a domain breadth axis, an activity axis, and an ability level axis. The system maps the domain breadth index, the mean of the active time trend chart, and the quantified ability level value to this coordinate system to form a closed curve of individual ability. Simultaneously, a preset job standard ability curve is superimposed on this coordinate system. The negative deviation area between the two curves is calculated, and an ability weakness label is generated based on the dimension with the largest deviation value, resulting in an ability profile dashboard.

[0011] Optionally, in the fourth implementation of the first aspect of the present invention, the test result data of the trainees is read, the difficulty coefficient in the job level quantification rule table is called, the scores of the correctly answered questions are summed by difficulty weighting, and the dynamic technical potential energy value is generated in the quantified ability level value. ,but: ; in, This represents the original score for the i-th correct question; Difficulty level; The kinetic slope of the most recent period of the active time series; This is a preset kinetic energy excitation factor.

[0012] Optionally, in the fifth implementation of the first aspect of the present invention, the technical supervision professional branch to which the trainee belongs is analyzed, and all personnel with the same target supervision field list are grouped into the same set to establish a homogeneous competition unit; the quantitative ability level values ​​of all personnel in the homogeneous competition unit are traversed, and the scores of the current trainee are sorted and percentiles are calculated to generate a field ranking index; the slope of the most recent time period of the active time trend graph is calculated to determine whether the interaction frequency of the trainee is in an accelerating growth trend or a declining trend, and a learning momentum slope is generated; the field ranking index is used as the vertical axis and the learning momentum slope is used as the horizontal axis, projected into a preset four-quadrant evaluation matrix, the quadrant type to which the trainee currently belongs is determined, and a dynamic incentive status label is generated; the dynamic incentive status label is read, and the public scope of the ability profile dashboard is controlled according to a preset transparency rule: if the label is a positive incentive type, a role model view visible to all members is generated and pushed to the public dashboard of the homogeneous competition unit; if the label is a negative type, a warning view visible only to the individual and the administrator is generated.

[0013] Optionally, in the sixth implementation of the first aspect of the present invention, the method includes: parsing the trainee's capability deficiency tag, reversing the business and standard association matrix, locating the specific technical supervision standard document identifier that caused the deficiency tag, and generating a list of standards to be remedied; traversing the trainee's historical error records within the list of standards to be remedied, extracting the standard original text clauses and analysis content corresponding to each error, and sorting them according to error frequency to generate a high-frequency error correction set; searching in the personnel database for other personnel whose quantitative capability level values ​​in the fields corresponding to the list of standards to be remedied reach a preset excellent threshold, marking them as candidate mentors, calculating the distance weight between the candidate mentors and the trainees in terms of geographical location or department affiliation, and selecting the one with the highest weight to generate a list of recommended support partners; using the list of standards to be remedied as a reading task, the high-frequency error correction set as a self-test task, and the list of recommended support partners as an offline interaction task, and encapsulating them according to a preset training cycle template to generate a personalized training plan.

[0014] Optionally, in the seventh implementation of the first aspect of the present invention, an isolation module is further included: physically dividing the hardware infrastructure running the NLP large model into a computing power domain, a storage domain, and a service access domain, and de-stacking the network architecture of each domain to form a physically isolated three-domain topology; in the three-domain topology, connecting the computing power domain and the storage domain to construct a 100G intranet high-speed channel that does not physically pass through the service access domain; configuring dual uplink Ethernet interfaces for the service access domain to establish a 25G external service interface independent of the computing power domain and the storage domain and facing the external user network; deploying the technical supervision standard data source in the storage domain and configuring access control policies to restrict the computing power domain to accessing data in the storage domain only through the 100G intranet high-speed channel, forming a secure and isolated operating environment.

[0015] Optionally, in the eighth implementation of the first aspect of the present invention, the secure and isolated operating environment is further configured with a management unit located in the management and scheduling service subnet. The management unit performs the following steps: receiving account information from the service platform through the 25G external service interface, synchronizing it with the local permissions within the secure and isolated operating environment, and generating a unified identity credential; intercepting operation requests sent by external users through the 25G external service interface, processing the operation requests using the unified identity credential, and generating a scheduling instruction; parsing the computing power requirements in the scheduling instruction, searching for idle GPU computing instances in the computing power domain, locking the specific hardware address that meets the requirements, and generating a target computing node index; sending a configuration command to the storage domain, instructing the storage domain to only grant RDMA access permissions to the hardware address corresponding to the target computing node index, and establishing a RoCE high-speed data transmission link that bypasses the management unit CPU and directly connects the storage domain and the computing power domain.

[0016] The mechanism of this invention is as follows: Orthogonal projection analysis is used to transform passive test-taking and active retrieval behaviors into a three-dimensional ability profile. Combined with a four-quadrant dynamic game and a complementary mentor matching mechanism, static training is transformed into a dynamic ecosystem with social incentive attributes. Beneficial effects: By using at least two different parsing algorithms to identify the original standard file in parallel, and by comparing multiple sets of parallel identified texts to determine the initial screening texts and marking items to be manually reviewed, the accuracy and efficiency of text parsing are greatly improved, and the problem of errors in traditional single parsing algorithms is effectively solved. Based on the basic information of the personnel, the standards and specifications that they should master in the specific power generation technology field are accurately matched to generate a customized question bank to meet the learning needs of different personnel, improve the pertinence and effectiveness of training, and avoid the "one-size-fits-all" training model. A comprehensive and in-depth assessment of trainees’ technical breadth, active time and ability level can accurately identify their skill gaps, providing a scientific basis for developing personalized training programs and helping trainees to improve their abilities in a targeted manner. The technical supervision professional branches to which the trainees belong are analyzed, and homogeneous competitive units are established. By calculating the field ranking index and learning momentum slope, dynamic incentive status labels are generated by projecting them onto a four-quadrant evaluation matrix. The scope of public disclosure of the capability profile dashboard is controlled according to preset transparency rules. Provide targeted learning content based on the personnel's skill gaps, including reading tasks, self-testing tasks, and offline interactive tasks, make full use of excellent internal resources to achieve precise training, help personnel quickly fill knowledge gaps, and improve professional skills; The hardware infrastructure for running large NLP models is physically divided into computing power domain, storage domain, and service access domain, forming a physically isolated three-domain topology. A 100G high-speed intranet channel and a 25G external service interface are built, and access control policies are configured. The management unit also realizes functions such as unified identity credentials, operation request processing, computing power scheduling, and high-speed data transmission link establishment. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an embodiment of the power industry training system based on NLP large model in this invention. Figure 2 This is a schematic diagram of another embodiment of the power industry training system based on NLP large model in this invention; Figure 3 A schematic diagram of the storage-computing separation architecture for the isolation module; Figure 4 This is a schematic diagram of an embodiment of a power industry training device based on a large NLP model, as described in this invention. Detailed Implementation

[0018] This invention provides a power industry training system based on a large NLP model, designed to offer personalized training solutions for the power industry. The terms "first," "second," "third," "fourth," etc. (if applicable) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0019] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the power industry training system based on NLP large models in this invention includes: 101. Data Module: Used to construct standard datasets in the field of technical supervision; Utilizes NLP large-scale models to parse, clean, and vectorize standard documents for multiple specialties and the entire production stage of thermal power plants, and establishes a vector database including technical supervision standards and specifications as a data source for technical supervision standards; It is understood that the executing entity of this invention can be a power industry training device based on a large NLP model, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0020] It should be noted that the standard documents were collected from authoritative sources in the thermal power industry, covering multiple professional disciplines such as thermal engines, electrical engineering, chemical monitoring, and environmental protection, as well as the entire production process from design and operation to maintenance. Specific data includes: Technical standards: including main plant layout requirements for the thermal engine section and lightning protection grounding design for the electrical section; Safety regulations: testing standards for components such as chloride ions and sulfates in water-soluble scale samples; Work instructions: on-site procedures such as boiler ignition operation procedures and turbine maintenance steps.

[0021] The collected documents include PDF, Word, and other formats. It is necessary to ensure the authority of the source (national standards, industry manuals), and initially classify and store them according to profession and production stage.

[0022] Using large NLP models to parse and clean unstructured text: Parsing key information: Extracting core entities ("boiler", "relay", "chloride ion content ≤0.01mg / g") from the document through Named Entity Recognition (NER), and extracting relationships between entities ("equipment-parameter", "standard-limit") through dependency parsing.

[0023] Cleaning and Standardization: Removing duplicate text (repeated common clauses across different documents) and irrelevant content (headers and footers). Standardizing terminology (unifying "TG-001" and "Steam Turbine 001" to "Steam Turbine TG-001"); handling missing values ​​(adding standard numbers or parameter units). Text standardization (unifying "≤10mg / L" to "less than or equal to 10 milligrams per liter"). After cleaning, the data is converted to a structured format, storing the fields "Standard Name - Specialty - Production Stage - Key Parameters" in tabular form.

[0024] The cleaned structured data was vectorized, and a vector database was built: Vectorization processing: The text was converted into vectors using an embedding model (BERT architecture): The "Annual Inspection Standard for Boilers and Pressure Vessels" was converted into a 384-dimensional vector. Technical terms ("supercritical units" and "flue gas desulfurization") were reduced in dimensionality and normalized to ensure semantic consistency of the vectors.

[0025] Constructing a vector database: Vectors are associated with metadata (standard source, effective date, applicable major) and stored in a vector database (Neo4j). Each vector record contains the following fields: [Standard ID: DL5461.1-2012, Vector: [0.12, -0.45, ...], Major: Thermomechanical, Production Stage: Design]. The database supports similarity retrieval (cosine similarity), facilitating rapid matching of relevant standards during subsequent question bank generation.

[0026] Quality Verification: Data reliability is ensured through both manual sampling verification (expert review of vector retrieval results for accuracy) and automatic evaluation (calculation of vector similarity threshold > 0.9). Regular Updates: The database is incrementally updated monthly based on standard revisions (DL / T series updates) or new operating procedures, and version records are maintained.

[0027] 102. Personnel Module: This module is used to obtain multi-dimensional basic attributes of trainees; it collects data on trainees' education, major, professional title, and work experience to form basic personnel information. It should be noted that the data sources are: academic certificates entered into the human resources system, verification data from the China Higher Education Student Information System (CHESICC), and employment registration forms. Example data: Educational level: Bachelor's degree (60%), Master's degree (30%), Associate's degree (10%). Major: Thermal Energy and Power Engineering (40%), Electrical Engineering (35%), Automation (15%), Chemical and Environmental Protection (10%). Institutional level: 985 / 211 universities (50%), regular undergraduate universities (30%), vocational colleges (20%).

[0028] Professional titles and professional qualifications. Data source: Enterprise professional title review records, National Professional Qualification Certification Database. Example data: Professional title levels: Junior Engineer (30%), Intermediate Engineer (45%), Senior Engineer (25%). Professional qualifications: Registered Electrical Engineer (15%), High Voltage Electrician Certificate (40%), Special Operation Certificate (30%). Review time: Before 2020 (20%), 2020-2025 (70%), After 2025 (10%).

[0029] Work experience and career trajectory. Data sources: Job transfer records in personnel files, project participation database, and performance appraisal system. Example data: Job type: Operation and maintenance (50%), overhaul and commissioning (30%), technical management (20%). Project experience: Participated in supercritical unit commissioning (25%), new energy grid connection projects (20%), and fault analysis projects (30%). Duration of work experience: Less than 5 years (35%), 5-10 years (40%), and more than 10 years (25%).

[0030] Skills certification and training records, data sources: internal corporate training system, external certification body data. Example data: Safety training: passed safety regulations exam (100%), emergency medical care certification (80%). Technical training: completed simulator training (60%), equipment fault diagnosis specialized training (40%).

[0031] The system automatically synchronizes education and professional title data from human resources systems (SAP, Oracle HR) via API interfaces, reducing manual data entry errors. For unstructured data (project reports, work summaries), NLP technology is used to extract key information, and named entity recognition is used to extract fields such as "participated in projects" and "responsible for equipment."

[0032] Deduplication and Completion: Merge multiple records for the same person (professional title promotion records), and infer missing fields (specialty) from historical project data. Terminology Standardization: Standardize different expressions such as "thermal power" and "energy and power engineering" into "thermal energy and power engineering".

[0033] Generate unique personnel IDs, link them to data from various dimensions, and form Table 1: Table 1 Dynamic updates: Monthly synchronization of personnel system change data (professional title promotion) to ensure real-time information.

[0034] 103. Standardization module, used to generate customized question banks; based on personnel basic information, it matches the standards and specifications that trainees should master in their specific power generation technology fields from the technical supervision standard data source, and generates customized question banks; It should be noted that the basic personnel information (from step 102) is as follows: Example data: Trainee A (Education: Master's degree; Major: Electrical Engineering; Title: Intermediate Engineer; Work experience: 5 years of substation operation and maintenance, participated in 3 relay protection renovation projects). Professional field tags: automatically resolved to "high voltage equipment maintenance", "relay protection", and "grid stability analysis".

[0035] Technical supervision standard data source (from step 101): Vector database content.

[0036] Multi-level matching strategy for professional and precise matching: Rule: Prioritize matching standards with high overlap with personnel's professional tags. For example, trainee A's "Relay Protection" tag will be directly associated with the clause in "DL / T 587-2016" regarding "protection setting error ≤ 5%". Matching data: Extract the top 10 most relevant standards (cosine similarity > 0.85) from the vector database.

[0037] Dynamic Difficulty Adjustment: Rule: The complexity of questions is adjusted based on professional title and years of experience. Intermediate engineers correspond to "Application Analysis" questions, while junior engineers correspond to "Basic Knowledge" questions. Example: Trainee A's 5 years of experience will trigger the generation of questions containing actual fault scenarios ("Analyze the causes of malfunction of 220kV line protection"), rather than basic concept questions. Timeliness Filtering: Rule: Only standards updated within the last 5 years (revised versions after 2020) are matched to ensure that the questions are consistent with current specifications.

[0038] Question generation and combination logic, question template design, multiple choice questions (60%): Example question: Question stem: "According to DL / T 587-2016, what should be the maximum operating time error of a relay protection device?" Options: A. 3ms B. 5ms C. 10ms D. 15ms (Correct answer: B). Generation logic: Key parameters are extracted from the standard as options; incorrect options are generated by adding or subtracting 20% ​​deviation.

[0039] Case Analysis Questions (30%): Example Question: Based on a real-world case of "maloperation of the main transformer differential protection" in a power plant, analyze which safety principle in GB / T 14285-2021 ("mismatch in current transformer selection") was violated. Generation Logic: Match similar cases from the historical fault database and associate them with the corresponding standard clauses.

[0040] Short answer questions (10%): Example question: "Briefly describe the requirements for the protection setting review cycle in DL / T 587-2016, and explain the risks of failing to review within the time limit." The number and distribution of questions are as follows: Each person has a fixed question bank of 50 questions, distributed according to the professional field weight: Main specialty (relay protection): 30 questions (20 multiple choice questions, 8 case analysis questions, and 2 short answer questions). Related specialty (high voltage equipment): 15 questions. General standards (safety regulations): 5 questions. Difficulty ratio: Basic questions 30%, intermediate questions 50%, and highly difficult questions 20%.

[0041] Automatic validation: Duplicate detection: Ensures that the similarity between questions within the same question bank is <15% (based on vector comparison). Answer consistency: Verifies that the answer is logically consistent with the standard original text through a rule engine.

[0042] Expert sampling review: 10% of the questions are randomly selected and scored by domain experts (accuracy requirement ≥95%). Problematic questions are returned for optimization. Output format: Question bank as shown in Table 2: Table 2 The attached standard original text excerpt is provided for your reference when answering the questions.

[0043] 104. Kanban module, used to build personnel competency profiles; obtain the test results of trainees on customized question banks, and combine the trainees' interaction data in daily question-and-answer scenarios to analyze the trainees' technical field breadth, active time and competency level, and generate a visual competency profile Kanban board and competency tags; It should be noted that the customized question bank test results are as follows: The test data recorded for trainees (using "Engineer Wang," an electrical engineer at a thermal power plant, as an example) are as follows: Question bank composition: 50 questions (30 electrical engineering questions, 15 thermal engine engineering questions, and 5 safety regulations questions). Test results: Electrical engineering accuracy rate 92% (errors concentrated on "relay protection setting"), thermal engine engineering accuracy rate 65% (70% error rate in the boiler system section), and safety regulations accuracy rate 100%. Test-taking behavior: Average test-taking time 45 seconds (30 seconds / question for electrical engineering, 70 seconds / question for thermal engine engineering), with a repeat error rate of 15%.

[0044] Daily Interaction Data: Q&A Scenario: In the past 30 days, Engineer Wang initiated 47 consultations on the training platform, including 35 electrical-related questions ("differential protection action logic"), 5 thermal engine-related questions ("steam turbine vibration standards"), and 7 interdisciplinary questions. Interaction Characteristics: The active period is weekdays from 14:00 to 17:00 (accounting for 70%), with an average dialogue depth of 3.2 rounds (4.5 rounds for electrical-related questions and only 1.5 rounds for thermal engine-related questions).

[0045] Remove invalid data (error records with a response time of less than 5 seconds) and standardize the timestamp format. Link the error points in the answers with the topic of the consultation questions (link the incorrect question "main transformer differential protection range" with the consultation record "differential protection CT selection").

[0046] Technical field breadth and professional coverage: The percentage of answers given by each specialty was calculated as follows: Electrical (60%), Thermal Engines (30%), and Safety (10%). Cross-domain correlation analysis: After consulting with Engineer Wang on thermal engine-related questions, the accuracy rate of related answers increased by 20% (after consulting on "Boiler and Pressure Vessel Standards," the accuracy rate for similar questions increased from 50% to 70%). Knowledge network construction: By extracting the co-occurrence relationships of keywords in Engineer Wang's interactive data through a topic model ("Relay Protection" is often associated with "CT Saturation" and "Setting Value Review"), a technical field map was generated, showing that his electrical knowledge was deep but his thermal engine knowledge was isolated.

[0047] Quantitative assessment of competency levels, tiered scoring: Basic cognition (0-30 points): Engineer Wang scored 28 points (full marks for safety regulations, only 15 points for basic thermal engine concepts). Application analysis (31-70 points): 62 points (85% accuracy in electrical fault diagnosis questions, but only 40% accuracy in thermal engine system analysis questions). Innovative decision-making (71-100 points): 45 points (lack of records of cross-disciplinary complex scenario deduction). Dynamic growth curve: Electrical professional competence has been stable over the past 3 months (fluctuation rate <5%), while thermal engine competence has improved by 20% through targeted learning (from 45 points to 65 points).

[0048] Cluster analysis of Wang's active time pattern shows that he is a "focused type in the afternoon": High-efficiency period: 14:00-16:00 (15% higher accuracy rate than average). Low-efficiency period: 09:00-11:00 (20% higher error rate, often accompanied by multiple short operations).

[0049] Visualized dashboards and tag generation, competency profile dashboard design, radar chart: displaying five-dimensional competency values ​​(professional breadth, depth, adaptability, activity level, and growth potential), with Engineer Wang displayed as "strong in electrical engineering, weak in thermal engineering". Trend chart: comparing monthly competency changes (the slope of the thermal engineering knowledge line is rising significantly). Heat map: marking the correlation between answer density and accuracy by hour (red high-error periods indicate avoiding complex tasks).

[0050] Ability Tagging System, automatically generating tags (based on threshold rules): Strength Tags: Electrical Protection Expert (Accuracy > 90%), Proficient in Safety Regulations (10 consecutive perfect scores). Weakness Tags: Thermal Engine Systems Beginner (Accuracy < 70%), Cross-Disciplinary Reasoning Needs Improvement (Cross-Domain Problem Solving Rate < 50%). Behavioral Tags: Highly Effective Afternoon Learner, Dependent on In-Depth Consulting. Tag Weights: Weights are assigned based on data confidence ("Electrical Protection Expert" weight 0.95, "Thermal Engine Systems Beginner" weight 0.88).

[0051] New answers and interaction data are synchronized weekly, and tags are recalculated (after Wang completed the thermal engine training, his tag was upgraded from "beginner" to "advanced"). Manual verification: Experts sample and review the accuracy of tags (10% of the "expert" tags are randomly selected to verify that their actual project experience matches ≥90%).

[0052] 105. The program module is used to generate personalized training programs; it identifies the trainees' skill gaps based on ability tags and intelligently generates personalized training programs that include targeted learning content.

[0053] It should be noted that the input data comes from the capability profile results in step 104, including capability tags (entry-level for thermal engine systems, expert-level for relay protection), capability score (score of 45 / 100 in the thermal engine field), and interaction behavior (average response time for thermal engine-related issues exceeds 50%).

[0054] Weakness Identification Rules: Threshold Judgment: Areas scoring below 70 points are automatically marked as weak points (65 points for thermal engine systems). Behavioral Correlation: Combining interaction data (the accuracy rate for thermal engine-related questions remains below 60%), the stability of the weak point is confirmed. Urgency Ranking: Based on job requirements (boiler maintenance personnel must prioritize thermal engine knowledge), weak points are divided into P0 (urgent), P1 (important), and P2 (general). For example, Engineer Wang (boiler maintenance personnel) has a weak point in thermal engine knowledge (P0) and a weak point in electrical knowledge (P2).

[0055] Knowledge structure deficiencies: Error clustering revealed that Engineer Wang's error rate on "Boiler and Pressure Vessel Standards" questions was as high as 70%, with errors concentrated in parameter memorization (design pressure limits, annual inspection cycles). Lack of practical skills: Historical project records show that Engineer Wang has not participated in boiler maintenance projects in the past year, indicating a lack of hands-on experience.

[0056] Training content matching and generation, intelligent matching of learning resources, matching logic: Extract standard clauses related to the shortcomings from the technical supervision standard database (step 101) (clauses regarding inspection cycles in "DL / T 561-2021 Boiler and Pressure Vessel Supervision Regulations"). Based on vector similarity (cosine similarity > 0.9) and difficulty adaptation (P0 shortcomings match "basic-application" level content), learning resources are selected.

[0057] Example matching results: Wang Gong's thermal engine weakness matches the following resources: Basic Theory: Boiler Structure and Principle Animation (15 minutes), Pressure Vessel Standard Interpretation Documents (3 articles). Application Case: Boiler Ignition Failure Fault Analysis Video (including interactive Q&A points). Simulation Training: Virtual Boiler Disassembly and Assembly Operation Module (requires 5 correct operations).

[0058] The learning path is dynamically designed with the following rules: Progressive learning: Content is arranged in the order of "basic concepts → case analysis → practical simulation" to ensure a reasonable cognitive gradient. Time adaptation: Complex content is pushed out during the active time period indicated in the competency profile (Wang Gong's most efficient time is 14:00-16:00).

[0059] Examples of generated schemes are shown in Table 3: Table 3 Behavioral feedback: If the simulated training fails 3 times in a row, additional tutoring materials (simplified operation guide) will be automatically inserted.

[0060] Short-term assessment: Weekly assessment scores are compared with competency profile data (once the warm-up score improves to 75 points, the learning intensity in that area is reduced). Long-term iteration: Monthly summaries of the effectiveness of all trainees' programs are conducted, and matching rules are optimized (if video format is found to be more effective for warm-up learning, resource priorities are adjusted).

[0061] Please see Figure 2 and Figure 3 Another embodiment of the power industry training system based on NLP large models in this invention includes: 201. Data Module: This module is used to construct a standard dataset for the field of technical supervision. It utilizes a large NLP model to parse, clean, and vectorize standard documents for multiple specialties and the entire production stage of thermal power plants, and establishes a vector database that includes technical supervision standards and specifications, serving as a data source for technical supervision standards. Specifically, the process involves: performing multi-algorithm composite verification; using at least two different parsing algorithms to perform parallel recognition of the original standard documents, outputting multiple sets of parallel recognized texts; comparing the multiple sets of parallel recognized texts, identifying the parts with consistent content as initial screening texts, and marking the parts with inconsistent content as items to be manually reviewed, which are then merged to form accurate text data; constructing a citation relationship graph; parsing the standard document numbers and cited clauses in the accurate text data to establish a standard citation relationship graph describing the mutual calling relationships between standard documents; performing timeliness and logical cleaning; traversing the standard citation relationship graph, checking whether the standard document numbers cited in the accurate text data exist in the graph nodes, removing text paragraphs that cite invalid nodes or non-existent nodes, and generating logically cleaned standard data; generating a vector index; segmenting and vector encoding the logically cleaned standard data to generate standard feature vectors, and storing the standard feature vectors in the database to form a technical supervision standard data source.

[0062] It should be noted that the construction of a standard library in the field of metal technical supervision of thermal power plants is taken as an example.

[0063] The system selected the "Technical Supervision Regulations for Metals in Thermal Power Plants" (DL / T 438-2016) as the original input file. Parallel recognition: The system simultaneously used both an "OCR algorithm based on layout analysis" and a "text extraction algorithm based on semantic understanding" to recognize the paragraph on page 56 of the document regarding "operating life of high-temperature headers." Comparison and verification: The first algorithm's recognition result was "when the operating time reaches 100,000 hours"; the second algorithm's recognition result was "when the operating time reaches 100,000 hours." The system comparison revealed an inconsistency between "100,000" and "100,000" in the text characters, marking it as requiring manual verification. Data merging: After manual verification confirmed that both were semantically identical, they were standardized to the Arabic numerals "100,000 hours" and merged with other consistent content in the document (which should undergo metallographic examination) to form accurate text data.

[0064] The system parsed the accurate text and identified that Clause 6.2 of DL / T 438-2016 explicitly referenced "Seamless Steel Tubes for High-Pressure Boilers" (GB / T 5310-2017). The system then created a directed edge in the graph database connecting the "DL / T 438-2016" node and the "GB / T 5310-2017" node, labeling the relationship as "reference clause," thus forming a standard reference relationship graph containing upstream and downstream calling logic.

[0065] The system iterates through the graphs and checks a section of the text data that references "Ultrasonic Testing of Welded Joints" (DL / T 820-2002). Graph verification reveals that this node is marked as "obsolete," and the currently valid version is 2019. The system determines that the reference in this section is invalid. To avoid misleading trainees, the system automatically removes 300 words of text containing this invalid reference, generating cleaned data.

[0066] The clause regarding "creep monitoring of main steam pipelines" was semantically segmented after cleaning, extracting the complete semantic block: "After 100,000 hours of operation, creep damage detection must be carried out on the main steam pipeline, with a detection rate of no less than 50% of the total number of bends." This text block was then converted into a high-dimensional numerical array (i.e., a standard feature vector) using a large-scale NLP model. This vector accurately captures core features such as "main steam pipeline," "100,000 hours," "creep damage," and "50%." This vector was stored in a vector database, completing the construction of the technical supervision standard data source.

[0067] 202. Personnel Module: This module is used to obtain multi-dimensional basic attributes of trainees; it collects data on trainees' education, major, professional title, and work experience to form basic personnel information. Specifically, a professional-job mapping table is established; a classification tree including all technical supervision process nodes in the power industry is pre-constructed, and each classification node is associated with specific professional background and work experience keywords to form a job-professional mapping table; target supervision areas are determined; the professional and work experience data of the collected trainees are input into the job-professional mapping table for keyword matching to filter out the technical supervision business branches corresponding to the trainees and generate a list of target supervision areas; a job level quantification rule table is constructed; a weight matrix based on professional title level and years of work experience is preset to assign corresponding initial difficulty coefficients to different combinations of professional titles and education levels to form a job level quantification rule table; an initial capability baseline is generated; the education and professional title data of the trainees are substituted into the job level quantification rule table to find the corresponding initial difficulty coefficients, and then bound to the target supervision area list to form basic personnel information including business scope constraints and difficulty parameter constraints.

[0068] It should be noted that the personnel file data of trainee "Li" was retrieved through the human resources interface. The specific attributes collected are: education level "Master of Energy and Power Engineering", major "Thermal Engineering", professional title "Senior Engineer", and work experience "12 years of experience in steam turbine maintenance and vibration fault diagnosis".

[0069] A "Job-Specialty Mapping Table" was pre-constructed. In the backend database, the category node "Steam Turbine Technical Supervision" was strongly associated with the keywords "Thermal Energy Engineering," "Rotating Machinery," and "Vibration Analysis"; the node "Insulation Technical Supervision" was associated with "Electrical Automation" and "High Voltage Testing." The system used Li's specialty "Thermal Energy Engineering" and his work experience "Vibration Fault Diagnosis" as input features and performed fuzzy keyword matching in the mapping table. The system successfully matched the business branches "Steam Turbine Technical Supervision" and "Vibration Technical Supervision," but not "Insulation Technical Supervision." Based on this, the system generated Li's target supervision area list as: [Steam Turbine Body Supervision, Vibration Supervision].

[0070] A pre-defined "Job Level Quantification Rule Table" was established. The rules are defined as follows: the basic difficulty weight for senior engineers is 0.7, for intermediate engineers it is 0.5; the correction coefficient for master's degree is +0.05; and the correction coefficient for 10 to 15 years of work experience is +0.1. The system substitutes Li's data into the rule table to calculate: the basic weight of 0.7 (senior engineer) plus the education correction of 0.05 (master's degree), plus the work experience correction of 0.1 (12 years of experience), yields Li's initial difficulty coefficient of 0.85. This coefficient represents his potential ability to handle highly complex technical problems.

[0071] The above calculation results were packaged to generate Li's "Basic Personnel Information" data package. This data package explicitly limited the scope of his training to "steam turbine and vibration monitoring" and locked the difficulty parameter to "0.85". This means that the subsequent test questions will focus on the more in-depth standard clauses in this field, rather than basic concepts, thus completing the transformation from raw files to a calculable baseline of personnel competence.

[0072] 203. Standards module, used to generate customized question banks; based on personnel basic information, it matches the standards and specifications that trainees should master in their specific power generation technology fields from the technical supervision standard data source, and generates customized question banks; Specifically, the process involves: constructing a business and standard association matrix; establishing a mapping relationship between all major process nodes of thermal power and technical supervision standard document identifiers based on the business logic of the entire power production process, forming a two-dimensional business and standard association matrix; locking in the scope of effective standards; parsing the target supervision area list in the personnel basic information, using this list to perform row index queries on the business and standard association matrix, extracting all standard document identifiers covered by the trainees' positions, and generating a list of effective standards; extracting compliance feature items; traversing the standard document content corresponding to the list of effective standards, using a pre-set strong constraint keyword table and numerical unit matching rules to selectively extract clauses containing mandatory descriptions such as "must," "strictly prohibited," and "should," or clauses containing specific physical quantity thresholds, generating a set of compliance feature items; synthesizing a pass / fail verification question bank; reading the initial difficulty coefficient in the personnel basic information, selecting the corresponding question type template based on the coefficient, filling the text descriptions in the compliance feature item set into the question stem area of ​​the template, and using the constraint conditions or parameter thresholds in the original clauses as standard answers to generate a customized question bank.

[0073] It should be noted that a two-dimensional matrix was established by analyzing the business logic of thermal power production. The vertical axis represents the specific nodes under the "boiler heating surface maintenance" business: "water-cooled wall tube bank inspection" and "superheater oxide scale cleaning"; the horizontal axis represents the standard document library identifier. The system marks the intersection of the business node "high-temperature superheater tube wall creep measurement" with the "Technical Supervision Regulations for Metals in Thermal Power Plants" (DL / T438-2016) as "strongly correlated" and the intersection with the "Electric Power Safety Work Regulations" as "weakly correlated," thus forming a digital mapping network.

[0074] The system reads the basic personnel information of trainee "Wang" and identifies his target supervision area as "boiler metal supervision". Using this tag, the system performs a row index query in the aforementioned matrix, filtering out irrelevant standards such as electrical and steam turbine standards, and accurately extracts a list of valid standards covered by Wang's position, including the core document DL / T 438-2016.

[0075] The system thoroughly analyzed the DL / T 438-2016 document in the valid list. Utilizing a table of strong constraint keywords such as "should" and "must," the system located Clause 6.3, which describes the expansion of alloy steel pipes. Simultaneously, using numerical unit matching rules, the system automatically extracted the specific physical quantity threshold from the clause: "For 12Cr1MoV alloy steel pipes, when the outer diameter creep expansion is greater than or equal to 2.5%, it should be replaced." The system extracted this text paragraph, containing both a logical judgment (greater than or equal to) and a handling action (replacement), as a compliance feature item.

[0076] The initial difficulty level for reading Wang's information was 0.6 (medium difficulty). Based on this, the system did not choose a simple fill-in-the-blank template, but instead used a "scenario-based judgment question" template. The system filled in the text from the feature entries into the question stem area and automatically generated the question: "During a major overhaul, the original outer diameter of a 12Cr1MoV high-temperature superheater tube was measured to be 50mm, and the current outer diameter is 51.3mm (expansion rate of 2.6%). Please determine how this tube should be handled?" Simultaneously, the system used "should be replaced" from the original clause as the standard answer and set "continue to observe" or "polish and repair" as distractors. This question was encapsulated in Wang's dedicated question bank to accurately verify his understanding of the core thresholds of metal monitoring.

[0077] 204. Kanban module, used to build personnel competency profiles; obtain the test results of trainees on customized question banks, and combine the trainees' interaction data in daily question-and-answer scenarios to analyze the trainees' technical field breadth, active time and competency level, and generate a visual competency profile Kanban and competency tags; Specifically, the process involves orthogonal classification of the interactive data; dividing the acquired data into test result data reflecting passive verification behavior and daily question-and-answer query logs reflecting active exploration behavior; parsing the daily question-and-answer query logs to extract technical keywords actively retrieved by trainees, forming a set of exploration intentions; calculating the technical field coverage; projecting the knowledge points of incorrect questions in the test result data and the keywords in the set of exploration intentions onto the classification tree structure of the technical supervision standard data source, calculating the ratio of the number of leaf nodes actually triggered by trainees to the total number of nodes in the target supervision field, and generating a technical field breadth index; constructing an activity time series; applying a sliding time window algorithm to the daily question-and-answer query logs to count the frequency of effective interactions within a unit time window, and marking the last interaction. The system generates an active time trend chart based on the timestamps of each effective interaction; it also generates a weighted competency matrix; it reads the test results data of trainees, calls the difficulty coefficients in the job level quantification rule table, and performs a difficulty-weighted summation of the scores for correctly answered questions to generate a quantified competency level value reflecting the trainee's actual technical depth; it draws a differentiated profile dashboard; it establishes a three-dimensional coordinate system including a domain breadth axis, an activity axis, and a competency level axis, and maps the domain breadth index, the mean of the active time trend chart, and the quantified competency level value onto this coordinate system to form a closed curve of individual competency; at the same time, it overlays a preset job standard competency curve on this coordinate system, calculates the negative deviation area between the two curves, generates competency weakness labels based on the dimension with the largest deviation value, and combines them to form a competency profile dashboard.

[0078] Furthermore, after generating a visualized capability profile dashboard and capability tags, the process also includes: constructing homogeneous competitive units; analyzing the technical supervision specialty branches to which trainees belong, grouping all personnel with the same target supervision field into the same set to establish homogeneous competitive units; calculating relative niches; iterating through the quantitative capability level values ​​of all personnel in the homogeneous competitive units, sorting and calculating the percentile of the current trainee's score, and generating a field ranking index reflecting their relative position in the group; analyzing momentum trends; calculating the slope of the most recent time period of the active time trend graph to determine whether the trainee's interaction frequency is increasing. Whether the trend is rapid growth or decline, a learning momentum slope is generated; an incentive state matrix is ​​synthesized; the domain ranking index is used as the vertical axis and the learning momentum slope is used as the horizontal axis, projected onto a preset four-quadrant evaluation matrix to determine the current quadrant type of the trainee and generate a dynamic incentive state label; a differentiated public view is generated; the dynamic incentive state label is read, and the public scope of the capability profile dashboard is controlled according to preset transparency rules: if the label is a positive incentive type, a role model view visible to all members is generated and pushed to the public dashboard of the homogeneous competition unit; if the label is a negative type, a warning view visible only to the individual and the administrator is generated.

[0079] It should be noted that the ability profile of "Engineer Zhang," a trainee in relay protection, is constructed as an example.

[0080] The system captures Zhang's behavioral data on the training platform in real time. Passive verification data: Zhang completed the "Relay Protection Setting Calculation" specialized test, answering 50 questions. Active exploration data: Logs show that Zhang actively searched for keywords such as "zero-sequence current transformer saturation" and "optical differential protection debugging" a total of 12 times in the daily question-and-answer box. The system orthogonally divides the above data; the former is used to calculate "hard skills," and the latter is used to analyze "exploration intentions."

[0081] The keywords triggered by Zhang Gong were projected onto the "Relay Protection Supervision" classification tree. This classification tree has 40 leaf nodes. Zhang Gong's incorrect answers covered 8 nodes, and the active retrieval covered 4 non-overlapping nodes, triggering a total of 12 nodes. The calculated technical field breadth index is 12 / 40 = 30%.

[0082] Active time trend: Using the 7-day sliding window algorithm, it was found that Zhang Gong's average daily interaction frequency in the past week was 15 times, and the last interaction occurred 2 hours before the current timestamp, generating an active curve with an upward trend.

[0083] Quantitative Ability Level: The system reads Zhang's answer data, in which he answered one question correctly (10 points) on "Differential Protection Malfunction Analysis" with a difficulty coefficient of 0.9, but also answered one basic concept question with a difficulty coefficient of 0.3 incorrectly. The system uses the following formula to calculate Zhang's dynamic technical potential energy value (E). tech It combines the depth of answering questions with learning motivation: Where: s i ω represents the original score (10 points) for the i-th correct question; i The difficulty coefficient (0.9) is derived from the "Job Level Quantification Rule Table"; k trend The kinetic slope of the most recent period of the active time series (calculated as 0.8 in this example, indicating an accelerated learning state); α is a preset kinetic energy activation factor (value 0.2). The arctan function is used to map the slope to a finite interval, preventing numerical distortion caused by short-term surges. Substituting the data, Zhang's basic ability score is 85 points; after kinetic correction, his quantitative ability level is 96.5 points.

[0084] A three-dimensional coordinate system was established, and Zhang's width (30%), average activity level (15), and ability value (96.5) were mapped to form a closed curve. After superimposing the standard curve for the "Senior Specialist" position, the system identified that Zhang had a negative deviation on the "width axis" and automatically generated a capability deficiency label of "narrow knowledge base".

[0085] A "homogeneous competition unit for relay protection" was constructed, comprising 15 professionals with similar expertise from across the plant. Engineer Zhang's score of 96.5 places him in the top 10%, indicating a high relative niche; and his kinetic slope k... trend > 0 (Positive Acceleration). The system projects this value into the "First Quadrant" (High Ability - High Growth) of the four-quadrant matrix, classifying it as a "Role Model" talent. Based on transparency rules, the system automatically generates a "Weekly Learning Star" role model view visible to all employees and pushes it to the department's public dashboard.

[0086] 205. The program module is used to generate personalized training programs; it identifies the trainees' skill gaps based on ability tags and intelligently generates personalized training programs that include targeted learning content.

[0087] Specifically, the process involves: tracing the shortcomings in standards; analyzing the trainees' capability deficiency tags, retrieving the business and standard correlation matrix, identifying the specific technical supervision standard document identifiers that led to the deficiency tags, and generating a list of standards to be remedied; extracting the mapping clauses for incorrect questions; traversing the trainees' historical incorrect question records within the list of standards to be remedied, extracting the corresponding original standard clauses and their explanations for each incorrect question, and sorting them by error frequency to generate a high-frequency incorrect question correction set; matching complementary mentors; searching the personnel database for other personnel whose quantitative capability level values ​​in the areas corresponding to the list of standards to be remedied reach a preset excellent threshold, marking them as candidate mentors; calculating the distance weight between candidate mentors and trainees in terms of geographical location or department affiliation, selecting the one with the highest weight to generate a list of recommended mentors; and synthesizing a composite training program; using the list of standards to be remedied as a reading task, the high-frequency incorrect question correction set as a self-test task, and the list of recommended mentors as an offline interactive task, packaging them according to a preset training cycle template to generate a personalized training program.

[0088] It should be noted that analyzing Liu's competency tags revealed his primary weakness as "insufficient ability to handle abnormal water and steam quality." The system used this tag to perform a reverse search of the "business-standard association matrix," precisely locating the core standard documents causing this weakness under the chemical supervision branch: the "Water and Steam Quality Standard for Thermal Power Plants" (GB / T 12145-2016) and the "Rust Prevention Guidelines for Shutdown (Standby) Thermal Equipment in Thermal Power Plants" (DL / T 956-2017). This generated a list of standards requiring remediation.

[0089] A review of Mr. Liu's historical answer records within the aforementioned standard range revealed that he made the most errors (a total of 5 errors) on questions related to "Dissolved Oxygen Control in Feedwater" and "Shutdown Protection." The system automatically extracted the original text from Chapter 4.2 of GB / T 12145-2016: "The dissolved oxygen content in feedwater of once-through boilers should be less than or equal to 7 micrograms per liter," along with the related analytical content: "Excessive dissolved oxygen will lead to electrochemical corrosion of the economizer and water-cooled walls." Based on the error frequency, the system prioritized these clauses, generating a set of corrections for frequently missed questions.

[0090] The system searches the personnel database for individuals with a quantitative competency score in chemistry exceeding 90. It identifies "Senior Engineer Zhao" (competency score 94) from the same department and "Engineer Li" (competency score 91) from another department. The system calculates distance weights: Engineer Zhao and Engineer Liu both belong to the "Chemistry subfield," so their geographical and organizational distance weight is 0.9; Engineer Li's weight is 0.6. Based on this, the system selects Senior Engineer Zhao as a recommended support partner and pushes his contact information to the plan.

[0091] The "one-week intensive sprint template" is used to encapsulate the extracted weaknesses criteria, incorrect question sets, and tutor information, generating a personalized training plan as shown in Table 4 below: Table 4 The plan was pushed to Engineer Liu's mobile learning interface, and the system will continuously track its completion progress.

[0092] 206. Isolation Module: Used to construct a storage-compute separation architecture; physically divides the hardware infrastructure running large NLP models into a computing domain, a storage domain, and a service access domain, and de-stacks the network architecture of each domain to form a physically isolated three-domain topology; establishes a low-latency dedicated channel; in the physically isolated three-domain topology, the RoCE network protocol is used to connect the computing domain and the storage domain, constructing a 100G intranet high-speed channel that does not physically pass through the service access domain; configures independent service links; configures dual uplink Ethernet interfaces for the service access domain, establishing a 25G external service interface independent of the computing domain and storage domain and facing the external user network; implements physical isolation of data assets; deploys technical supervision standard data sources in the storage domain and configures access control policies to restrict the computing domain to accessing data in the storage domain only through the 100G intranet high-speed channel, forming a secure and isolated operating environment.

[0093] Specifically, the secure and isolated operating environment is also equipped with a management unit located in the management and scheduling service subnet. This management unit performs the following steps: establishing identity and permission mapping; receiving account information from the service platform via the 25G external service interface, synchronizing it with local permissions within the secure and isolated operating environment, and generating unified identity credentials including access levels and resource quotas; performing out-of-band instruction cleansing; intercepting operation requests sent by external users through the 25G external service interface, using the unified identity credentials to perform legality verification and desensitization processing on the operation requests, and generating cleaned scheduling instructions; locking computing resource nodes; parsing the computing power requirements in the cleaned scheduling instructions, searching for idle GPU computing instances in the computing power domain, locking specific hardware addresses that meet the requirements, and generating a target computing node index; constructing a zero-copy data channel; sending a configuration command to the storage domain, instructing the storage domain to only grant RDMA access permissions to the hardware address corresponding to the target computing node index, and establishing a RoCE high-speed data transmission link that bypasses the management unit CPU and directly connects the storage domain and the computing power domain.

[0094] It should be noted that the data center racks are physically divided into three independent zones. The computing zone deploys eight high-performance computing nodes (configured with NVIDIA A800 GPUs) for running inference computations on large NLP models. The storage zone deploys an all-flash distributed storage cluster, storing 500TB of vectorized technical supervision standard data sources. The service access zone deploys a web server cluster, directly serving enterprise intranet users. The three zones are no longer connected using traditional hierarchical stacked networks, but rather with physically isolated cabling.

[0095] The specific network parameter configurations are shown in Table 5 below, ensuring complete separation between data flow and control flow: Table 5 When a trainee initiates a request to "query boiler water-cooled wall maintenance standards" through a browser, the management unit (a stand-alone scheduling server) located in the management subnet performs the following operations: Establish identity and permission mapping: The management unit receives the account "User_9527 (Zhang Gong)" from the business middle platform through the 25G external interface. The system immediately queries the local permission table, maps it to the internal temporary credential "Token_0X1A", and grants the credential "read-only reasoning permission" to the technical supervision knowledge base, strictly blocking its permission to modify or download the underlying data.

[0096] Out-of-band instruction cleaning: The system intercepts the original request text, performs regular expression filtering, removes special characters that may lead to SQL injection, and standardizes the request into the cleaned scheduling instruction: "[Action: Query] [Target:Water_Wall_Maintenance]".

[0097] Locking the computing resource node: The management unit scans the computing domain status and finds that the current GPU utilization of "Compute Node_04" is 12%, indicating it is in an idle state. The system immediately locks the physical address of this node (MAC address: AA:BB:CC:DD:EE:04) and generates a target index.

[0098] Establishing a zero-copy data path: This is the most crucial step. The management unit sends control commands to the storage domain, instructing it to enable the "RDMA (Direct Memory Access)" switch. The storage domain then establishes a direct data path that bypasses both the management unit's CPU and the business domain. Standard feature vector data stored in the SSD is directly written to the video memory of "Compute Node_04" via DMA (Direct Memory Access) through the 100G RoCE intranet, without any CPU copy operations. Data transfer latency is controlled at the microsecond level, thus achieving ultimate inference response in a physically isolated environment.

[0099] The above describes the NLP-based large-scale power industry training device in the embodiments of the present invention in detail from the perspective of modular functional entities. The following describes the NLP-based large-scale power industry training device in the embodiments of the present invention in detail from the perspective of hardware processing.

[0100] Figure 4 This is a schematic diagram of the structure of a power industry training device based on a large NLP model, provided in an embodiment of the present invention. The power industry training device 300 based on the large NLP model may include a processor 301 and a memory 302. The memory 302 is used to store program instructions and / or data, and the processor 301 is used to execute the program instructions stored in the memory 302, thereby implementing the method described in the above embodiment.

[0101] Optionally, the memory 302 and the processor 301 are coupled. The coupling is an indirect coupling or communication connection between devices, units, or modules, and can be electrical, mechanical, or other forms, for information interaction between devices, units, or modules.

[0102] Optionally, the NLP-based large-scale power industry training device 300 may also include a communication interface 303. The communication interface 303 is used to communicate with other devices via a transmission medium, for example, transmitting received signals from other communication devices to the processor 301, or transmitting signals from the processor 301 to other communication devices. The communication interface 303 can be a transceiver or an interface circuit, such as a transceiver circuit or transceiver chip.

[0103] This application embodiment does not limit the specific connection medium between the processor 301, memory 302, and communication interface 303. This application embodiment... Figure 4 The processor 301, memory 302, and communication interface 303 are connected via a bus 304. Figure 4 The connections between other components are shown in bold and are for illustrative purposes only, not as limiting information. The bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0104] The present invention also provides a power industry training device based on an NLP large model. The power industry training device based on an NLP large model includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of the power industry training system based on an NLP large model in the above embodiments.

[0105] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the power industry training system based on the NLP large model.

[0106] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0107] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

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

Claims

1. A power industry training system based on NLP large-scale models, characterized in that, include: The data module is used to process standard documents for multiple disciplines and the entire production stage of thermal power using large NLP models, establish a vector database, and obtain technical supervision standard data sources. The personnel module is used to collect data on trainees' education, major, professional title, and work experience to form basic personnel information; The standardization module is used to match the standards and specifications that trainees should master in the specific field of power generation technology based on the basic personnel information and the technical supervision standard data source, and generate a customized question bank. The Kanban module is used to obtain the test results of trainees on the customized question bank, and combine the trainees' interaction data in daily question-and-answer scenarios to analyze the trainees' technical field breadth, active time and ability level, and generate ability profile Kanban and ability tags. The solution module is used to identify the trainees' skill gaps based on the skill tags and generate personalized training solutions that include targeted learning content.

2. The power industry training system based on NLP large model according to claim 1, characterized in that, include: At least two different parsing algorithms are used to perform parallel recognition on the original standard file, and multiple sets of parallel recognition texts are output. The multiple sets of parallel recognition texts are compared, and the parts with consistent content are identified as the initial screening texts, while the parts with inconsistent content are marked as items to be manually reviewed. After review, they are merged to form accurate text data. Parse the standard document numbers and referenced clauses in the accurate text data to establish a standard reference relationship map describing the mutual calling relationships between standard documents; Traverse the standard reference relationship graph, check whether the standard document number referenced in the accurate text data exists in the graph node, remove text paragraphs that reference invalid nodes or non-existent nodes, and generate logically cleaned standard data. Standard data is segmented and vector-encoded to generate standard feature vectors, which are then stored in a database to form a standard data source for technical supervision.

3. The power industry training system based on NLP large model according to claim 2, characterized in that, include: A classification tree is pre-constructed that includes all technical supervision process nodes in the power industry, and each classification node is associated with keywords to form a job-professional mapping table; The collected data on the trainees' professional background and work experience are input into the job-professional mapping table for keyword matching, and the technical supervision business branches corresponding to the trainees are filtered out to generate a list of target supervision areas. A weight matrix based on professional title level and years of work experience is preset to assign corresponding initial difficulty coefficients to different combinations of professional titles and educational backgrounds, forming a job level quantification rule table; The trainees' educational background and professional title data are substituted into the job level quantification rule table to find the corresponding initial difficulty coefficient, and then bound to the target supervision area list to form the basic personnel information.

4. The power industry training system based on NLP large model according to claim 3, characterized in that, include: The acquired data is divided into test result data reflecting passive verification behavior and daily question and answer query logs reflecting active exploration behavior. The daily question and answer query logs are analyzed to extract the technical keywords actively searched by the trainees and form a set of exploration intentions. The knowledge points to which the wrong questions in the test results data belong and the keywords in the set of exploration intentions are projected onto the classification tree structure of the technical supervision standard data source. The ratio of the number of leaf nodes actually triggered by the trainees to the total number of nodes in the target supervision field is calculated to generate a technical field breadth index. The sliding time window algorithm is applied to the daily question and answer query logs to count the frequency of effective interactions within a unit time window, and the timestamp of the last effective interaction is marked to generate an active time trend chart. Read the test results data of the trainees, call the difficulty coefficient in the job level quantification rule table, perform difficulty-weighted summation on the scores of the correct answers, and generate a quantitative ability level value; A three-dimensional coordinate system is established, including a domain breadth axis, an activity level axis, and a capability level axis. The domain breadth index, the mean of the activity time trend chart, and the quantified capability level value are mapped onto this coordinate system to form a closed curve of personal capability. At the same time, a preset job standard capability curve is superimposed on this coordinate system. The negative deviation area between the two curves is calculated, and capability weakness labels are generated based on the dimension with the largest deviation value to obtain a capability profile dashboard.

5. The power industry training system based on NLP large model according to claim 4, characterized in that, The test results data of the trainees are read, the difficulty coefficients in the job level quantification rule table are called, the scores of the correctly answered questions are summed with difficulty weighting, and the dynamic technical potential value is generated in the quantitative ability level value. ,but: ； in, This represents the original score for the i-th correct question; Difficulty level; The kinetic slope of the most recent period of the active time series; This is a preset kinetic energy excitation factor.

6. The power industry training system based on NLP large model according to claim 4, characterized in that, Analyze the technical supervision specialty of the trainees, group all personnel with the same target supervision field into the same set, and establish homogeneous competitive units; Iterate through the quantitative ability level values ​​of all personnel in the homogeneous competitive unit, sort the scores of the current trainees and calculate the percentiles to generate a domain ranking index; The slope of the most recent time period of the active time trend graph is calculated to determine whether the interaction frequency of the trainees is in an accelerating growth trend or a declining trend, and a learning momentum slope is generated. Using the domain ranking index as the vertical axis and the learning momentum slope as the horizontal axis, the data is projected onto a preset four-quadrant evaluation matrix to determine the current quadrant type of the trainee and generate a dynamic incentive status label. The dynamic incentive status label is read, and the public scope of the capability profile dashboard is controlled according to the preset transparency rules: if the label is a positive incentive type, a role model view that is visible to all members is generated and pushed to the public dashboard of the homogeneous competition unit; if the label is a negative type, an early warning view that is visible only to the individual and the administrator is generated.

7. The power industry training system based on NLP large model according to claim 4, characterized in that, include: The system analyzes the skill gap labels of trainees, reversely retrieves the business and standard association matrix, locates the specific technical supervision standard document identifier that caused the gap label, and generates a list of standards to be remedied. The system iterates through the trainees' historical error records within the list of standards to be remedied, extracts the original standard clauses and explanations corresponding to each error, sorts them by error frequency, and generates a high-frequency error correction set. Search the personnel database for other personnel whose quantitative ability level value reaches the preset excellent threshold in the field corresponding to the list of standards to be remedied, mark them as candidate mentors, calculate the distance weight between the candidate mentor and the trainee in terms of geographical location or department affiliation, and select the one with the highest weight to generate a list of recommended support partners. The list of standards to be remedied is used as a reading task, the set of frequently asked wrong questions is used as a self-test task, and the list of recommended support partners is used as an offline interactive task. These are all packaged according to a preset training cycle template to generate a personalized training program.

8. The power industry training system based on NLP large model according to claim 1, characterized in that, It also includes an isolation module: The hardware infrastructure running the large NLP model is physically divided into a computing power domain, a storage domain, and a service access domain. The network architecture of each domain is de-stacked to form a physically isolated three-domain topology. In the three-domain topology, the computing power domain and the storage domain are connected to construct a 100G high-speed intranet channel that does not physically pass through the service access domain. Configure dual uplink Ethernet interfaces for the service access domain to establish a 25G external service interface that is independent of the computing power domain and storage domain and faces the external user network. The technical supervision standard data source is deployed in the storage domain, and an access control policy is configured to restrict the computing power domain to accessing the data in the storage domain only through the 100G intranet high-speed channel, thus forming a secure and isolated operating environment.

9. The power industry training system based on NLP large model according to claim 8, characterized in that, The secure, isolated operating environment is also configured with a management unit located in the management and scheduling service subnet, through which the following steps are performed: The account information of the business platform is received through the 25G external business interface, and synchronized with the local permissions in the secure and isolated operating environment to generate a unified identity credential. Intercept operation requests sent by external users through the 25G external service interface, process the operation requests using the unified identity credential, and generate scheduling instructions; The computing power requirement in the scheduling instruction is parsed, an idle GPU computing instance is searched in the computing power domain, a specific hardware address that meets the requirement is locked, and a target computing node index is generated. A configuration command is sent to the storage domain, instructing the storage domain to grant RDMA access permissions only to the hardware address corresponding to the target computing node index, thereby establishing a RoCE high-speed data transmission link that bypasses the management unit CPU and directly connects the storage domain and the computing domain.

Images