Thermal power plant operation analysis method and system based on AI large model

Abstract

This invention relates to the field of industrial data analysis technology, and discloses a method and system for analyzing the operation of thermal power plants based on an AI large-scale model. The method includes: calling an inference model to perform intent parsing on natural language requests and initializing the task context; generating query statements to obtain structured operational data and performing anomaly distribution checks; for data that passes the checks, generating program code to execute in an isolated sandbox and outputting anomaly indicators; recalling domain rule text based on the anomaly indicators, performing fault association mapping to output diagnostic text; extracting numerical characters from the diagnostic text, performing tolerance matching verification with the structured operational data to generate factual consistency verification results; and finally integrating and outputting a thermal power plant operation diagnostic report. This invention eliminates noise interference from the underlying field data, intercepts false values ​​generated by model inference, and ensures that the output industrial diagnostic conclusions have solid factual basis.

Description

Technical Field

[0001] This invention relates to the field of industrial data analysis technology, specifically to a method and system for analyzing the operation of thermal power plants based on AI large-scale models. Background Technology

[0002] With the development of artificial intelligence technology, large language models are gradually being introduced into the operation analysis of thermal power plants, aiming to assist in equipment status diagnosis and data query through natural language interaction. However, in actual industrial applications, the low-level time-series data collected by field sensors is often accompanied by missing, frozen, and abrupt noise. Directly processing this raw, heterogeneous data without cleaning by large language models can lead to serious logical biases at the input end. When dealing with complex operating conditions and performing multi-step logical reasoning and external tool scheduling, the model often generates redundant operation instructions with different expressions but semantically equivalent meanings, causing the system to fall into an invalid reasoning loop and causing the analysis task to stall.

[0003] At the same time, due to the inherent limitations of the autoregressive generation mechanism of large language models, when outputting long text diagnostic conclusions, the model is prone to generating false quantitative index values ​​out of thin air, detached from objective time-series sensor data. This results in the final industrial analysis report lacking solid underlying factual support, making it difficult to meet the safety and reliability requirements of serious industrial scenarios such as thermal power plants. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for analyzing the operation of thermal power plants based on AI large models. This solves the problems that existing large language models, when applied to the analysis of thermal power plant operation, are susceptible to interference from abnormal underlying data, leading to inference bias; are prone to getting stuck in invalid tool call loops during multi-step inference; and generate diagnostic reports that are detached from objective facts and lack underlying basis.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a method for analyzing the operation of thermal power plants based on a large AI model, comprising the following steps: The encoding and decision reasoning model is invoked to perform intent parsing on the received natural language request, extract the device object, time interval, and analyze target indicators to complete the task context initialization; The system calls a dedicated text-to-SQL model and combines it with the database schema to generate a query statement. The query statement is then executed to obtain structured running data, and anomaly distribution checks are performed on the structured running data. For the structured runtime data that has passed the anomaly distribution test, the encoding and decision reasoning model is invoked to generate program code and execute it in an isolated execution sandbox, outputting data frames, chart files and anomaly indicators; Based on the abnormal indicators, a diagnostic question instruction is generated and the domain rule text is recalled. The domain rule text and the abnormal indicators are input into the domain knowledge diagnostic model to perform fault association mapping and output the diagnostic text. Numerical characters are extracted from the diagnostic text, and tolerance matching verification is performed between the numerical characters and the structured runtime data to generate a fact consistency verification result. Based on the factual consistency verification results, the data frame, the chart file, and the diagnostic text, a thermal power plant operation diagnostic report is output.

[0006] In one specific implementation, performing anomaly detection on the structured runtime data includes: The system identifies missing values, frozen values, and abrupt outliers in the structured operational data, and calculates the percentage of outliers in the total sample data. When the percentage of outliers exceeds a preset 10% threshold, a basic quality report is output. When the percentage of outliers is not greater than the preset 10% threshold, interpolation repair and smoothing denoising are performed on the structured operational data, and a maximum-minimum mapping operation is performed to obtain the structured operational data that passes the outlier distribution test. The abrupt outliers are determined by calculating the absolute deviation of the local data point from the arithmetic mean according to the Laida criterion, where the difference is greater than 3 times the standard deviation.

[0007] Through the aforementioned data quality assessment and cleaning mechanisms, this invention achieves proactive interception of common data degradation and distortion in industrial settings, utilizes the Laida criterion to filter out noise interference that deviates from the normal physical fluctuation range, and eliminates the dimensional differences between multi-source heterogeneous sensors, providing a reliable data benchmark for large-scale model inference.

[0008] In one specific implementation, during the process of calling the encoding and decision reasoning model to generate program code and executing it in an isolated execution sandbox, a duplicate call interception mechanism is further included: extracting the parameter word set of the current call to the encoding and decision reasoning model action and the parameter word set of historical call records in the same session thread; performing a basic comparison between the parameter word set of the current call to the encoding and decision reasoning model action and the parameter word set of the historical call records using a hash algorithm; when the comparison results are inconsistent, calculating the number of words contained in the intersection of the parameter word set of the current call to the encoding and decision reasoning model action and the parameter word set of the historical call records, and the number of words contained in the union, dividing the number of words contained in the intersection by the number of words contained in the union to obtain a semantic similarity coefficient; when the semantic similarity coefficient is greater than or equal to a preset interception threshold of 0.55, or when the cumulative number of calls of a single dedicated execution tool reaches a preset maximum call limit, the call action is intercepted and the current reasoning branch is terminated.

[0009] The aforementioned multi-layered repeated call interception mechanism based on the Jaccard coefficient can quantify the degree of semantic overlap between action instructions and business requirements, thereby cutting off redundant requests with different textual expressions but equivalent semantics and improving the efficiency of multi-step reasoning.

[0010] In one specific implementation, the step of generating diagnostic query instructions and recalling domain rule text based on the abnormal indicators includes: The pre-acquired external specification documents are divided into different professional categories and processed by slicing and block processing. The discrete text blocks are mapped into high-dimensional dense vectors and attached with professional category labels for persistent storage in a vector database. Diagnostic question instructions containing target category labels are generated based on the anomaly indicators. The target category labels in the diagnostic question instructions are extracted and metadata filtering is performed on the vector database. An approximate nearest neighbor retrieval algorithm is used to obtain preliminary candidate texts and semantic similarity spatial distance scores based on Euclidean distance. Key terms in the diagnostic question instructions are traversed, and the frequency of occurrence of the key terms in the preliminary candidate texts and the inverse document frequency in the entire knowledge base corpus are counted to calculate the matching weight coefficient. The priority order of the preliminary candidate texts is rearranged by combining the semantic similarity spatial distance score and the matching weight coefficient, and the domain rule text is output.

[0011] By integrating semantic retrieval based on Euclidean distance with keyword retrieval that assesses word density, this mechanism effectively reduces the retrieval computation space and lowers the hit rate of proper nouns in the thermal power industry.

[0012] In one specific implementation, the step of extracting numerical characters from the diagnostic text, performing tolerance matching verification between the numerical characters and the structured runtime data, and generating a factual consistency verification result includes: The system uses regular expression matching logic to extract numeric characters and local text from the diagnostic text as context information, converting the numeric characters into discrete numeric variables. It then iterates through the rows and columns of the structured data, extracting original numeric fields to establish a candidate matching set. The system calculates the absolute value of the difference between the discrete numeric variables and the original numeric fields in the candidate matching set, dividing the absolute value of the difference by the absolute value of the original numeric fields to obtain the relative error. If the relative error is less than 1%, the verification is considered successful; if the relative error is not less than 1%, a warning flag is added to the corresponding unmatched position in the diagnostic text, generating the factual consistency verification result.

[0013] In one specific implementation, the steps of extracting numerical characters from the diagnostic text, performing tolerance matching verification between the numerical characters and the structured runtime data, and generating a factual consistency verification result further include: The diagnostic text contains derived index values ​​resulting from complex mathematical operations; the generation results of the derived index values ​​and the basic structured data on which the operations depend are extracted and input into the coding and decision reasoning model to perform reverse mathematical calculations; when the final result value obtained by the reverse mathematical calculations deviates from the derived index values ​​output in the diagnostic text, a warning label is added to the derived index values.

[0014] This fact consistency verification mechanism transforms quantitative descriptions in unstructured text into mathematically verifiable variable comparisons through underlying data tracing and relative error calculation. While being compatible with the mantissa deviations of conventional floating-point arithmetic, it intercepts and eliminates abnormal values ​​generated by model calculation errors.

[0015] In one specific implementation, after the task context initialization is completed, the method further includes: Monitor the number of lexical units in the current session history corresponding to the task context, and calculate the ratio of the number of lexical units to the maximum capacity of the context window of the encoding and decision reasoning model; when the ratio reaches or exceeds 90%, activate the context compression module; retain the original interaction session of the most recent interaction round, and call the encoding and decision reasoning model to generate a structured summary of the earlier historical session; concatenate the structured summary with the retained original interaction session and re-inject it into the context of the encoding and decision reasoning model.

[0016] In one specific implementation, the analysis method further includes an execution state recovery mechanism, comprising: The background monitoring component monitors the execution status of the encoding and decision reasoning model in real time. When an execution stall is detected, such as an empty response exception, a transitional statement exception, or an unexpected termination of streaming execution, the current regular reasoning process is paused. The core summary of the execution results of existing external tools is extracted from the recorded current dialogue history and the large model call trajectory log. The core summary and the task intent obtained by the intent parsing are assembled into a context prompt word and re-injected into the current execution thread.

[0017] In one specific implementation, the analysis method further includes executing a backup degradation strategy, including: When the number of failed attempts to re-inject the currently executing thread reaches the preset maximum number of recovery attempts, the system cache pool is accessed to collect temporarily stored query statements, underlying return results, program code analysis output, domain knowledge diagnostic conclusions, and partial results of the chart file list. The collected partial results are assembled according to a preset standard markup language formatting template, and the degraded state power plant operation diagnostic report is encapsulated and output.

[0018] The aforementioned dynamic context compression, state recovery, and backup degradation strategies enable real-time monitoring of physical computing resource limitations and model inference breakage, overcome the risk of context truncation under limited video memory constraints, and force the output of backup basic results to end the task flow in abnormal stagnation states.

[0019] A second aspect of this invention provides a thermal power plant operation analysis system based on an AI large-scale model, comprising: an intelligent agent orchestration and control module, used to perform the following operations: The system invokes an encoding and decision-making reasoning model to parse the intent of received natural language requests, extracting device objects, time intervals, and target indicators to complete task context initialization. It then invokes a text-to-SQL model combined with a database schema to generate query statements. For the structured operational data that has passed anomaly distribution testing, the system invokes the encoding and decision-making reasoning model to generate program code, which is then executed in an isolated execution sandbox, outputting data frames, chart files, and anomaly indicators. Based on the anomaly indicators, diagnostic query instructions are generated. Finally, based on the factual consistency verification results, the data frames, the chart files, and the diagnostic text, a thermal power plant operation diagnostic report is output. The multidimensional operational data acquisition module is used to run the query statement to obtain structured operational data and perform anomaly distribution checks on the structured operational data. The knowledge base retrieval module is used to recall domain rule texts based on the diagnostic question instructions; A dedicated model support module is used to provide the encoding and decision reasoning model, the text-to-SQL dedicated model, and to perform fault association mapping by inputting the domain rule text and the anomaly index into the domain knowledge diagnosis model and outputting the diagnosis text; The degradation-tolerance control module is used to extract numerical characters from the diagnostic text, perform tolerance matching verification between the numerical characters and the structured running data, and generate the factual consistency verification result.

[0020] This invention provides a method and system for analyzing the operation of thermal power plants based on a large AI model. It offers the following advantages: 1. This invention performs anomaly distribution testing on the acquired structured operational data, identifies abrupt outliers using the Laida criterion, and performs interpolation repair, smoothing and denoising, and maximum-minimum mapping operations on the data. This eliminates noise data generated by on-site sensors in thermal power plants due to physical degradation or electromagnetic interference, eliminates the dimensional scale differences of multi-source heterogeneous data, and avoids inference bias caused by large language models directly reading abnormal underlying data, providing an objective and reliable data benchmark for subsequent in-depth analysis and calculation.

[0021] 2. This invention introduces a duplicate call interception mechanism in the process of model generation program code and scheduling execution tool. By extracting the parameter lexical set, it sequentially performs hash-based comparison and semantic similarity calculation based on Jaccard coefficient. It can quantify the equivalence of substantive business requirements between different textual expressions, effectively intercept redundant call instructions generated by the large language model in multi-step inference, block the path of the system falling into invalid inference loop, and improve the overall flow efficiency of complex working condition analysis tasks.

[0022] 3. This invention establishes a fact consistency verification mechanism based on underlying data tracing. It calls regular expressions to extract numerical characters from diagnostic text and performs tolerance matching verification with the original structured running data. Under the premise of being compatible with the common deviations of computer floating-point arithmetic, it transforms the quantitative description of unstructured text into variable comparison, intercepts fabricated values ​​that deviate from objective facts due to the limitations of the underlying generation mechanism of the model, and ensures that the final output running diagnostic report has solid factual basis. Attached Figure Description

[0023] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a flowchart of the deep analysis calculation and visualization rendering process of the present invention; Figure 4 This is a flowchart of the factual consistency verification process of the present invention; Figure 5 This is a flowchart of the output and degradation response control of the present invention; Figure 6 This is a comparison chart of the data quality assessment and smoothing / denoising effects of the present invention. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See attached document Figure 1 This application provides a thermal power plant operation analysis system based on a large language model. The system includes: a multi-dimensional operation data acquisition module, a dedicated model support module, a knowledge base retrieval module, an intelligent agent orchestration and control module, and a degradation and fault tolerance control module.

[0026] The multi-dimensional operational data acquisition module is used to access heterogeneous data sources from the main and auxiliary equipment of a thermal power plant. This module connects to the distributed control system or plant-level monitoring information system to acquire time-series sensor parameters, connects to the operation and maintenance log system to obtain ledger reports, and receives offline characteristic parameters input from the monitoring equipment. The multi-dimensional operational data acquisition module has a built-in data cleaning unit that identifies missing values, frozen values, and abrupt changes in the accessed underlying time-series data, and performs data interpolation, repair, smoothing, noise reduction, and standardization.

[0027] The dedicated model support module serves as the core of the system's reasoning, integrating three independent large language models fine-tuned using low-rank adaptive instructions. Specifically, these include a domain knowledge diagnostic model for handling domain knowledge reasoning and equipment fault attribution, a text-to-SQL dedicated model for converting natural language into structured query language, and an encoding and decision-making reasoning model for driving task planning and code generation.

[0028] The knowledge base retrieval module stores and retrieves external specification documents and case studies related to power plant operation. This module slices operating procedures and maintenance log texts and converts them into high-dimensional vectors for persistent storage. When executing a retrieval request, the knowledge base retrieval module employs a mechanism combining semantic approximate nearest neighbor retrieval and keyword retrieval to retrieve external factual reference texts that match the current diagnostic task.

[0029] The agent orchestration and control module is built upon the inference-action paradigm. This module allocates dedicated execution toolkits to agents, including query execution, code execution, and graph rendering. It incorporates an isolated execution sandbox, placing the invoked program code in an isolated environment with configured memory and computing resource limits, and provides a cross-execution step context variable sharing mechanism to maintain system state.

[0030] The degradation and fault tolerance control module monitors the system's operational status and ensures output consistency. This module performs context compression by monitoring the computational resource consumption of the dialogue history, performs numerical cross-validation on the analysis conclusions to verify factual consistency, and triggers a backup degradation response when the agent's inference stalls.

[0031] See attached document Figure 2 This application provides a method for analyzing the operation of thermal power plants based on a large language model, comprising the following steps: S10: Receive the operation analysis request instruction and initialize the task context. The intelligent agent orchestration and control module receives the natural language request input by the operator, calls the encoding and decision reasoning model to parse the intent of the input request, and decomposes it into the device objects, time intervals, and analysis target indicators involved in the current request; S20, Execution of Data Query and Quality Assessment. The intelligent agent orchestration and control module calls a dedicated text-to-SQL model and combines it with the database schema to generate query statements. Subsequently, it calls an isolated execution sandbox to run the query statements and obtain structured runtime data from the underlying database. The multi-dimensional runtime data acquisition module performs anomaly detection on the acquired structured runtime data to determine data reliability. S30 performs in-depth analysis calculations and visualization rendering. For data that has passed the quality assessment, the agent orchestration and control module again calls the encoding and decision reasoning model to generate program code for statistical calculations and visualization. The program code is then passed to the isolated execution sandbox for decoding and execution, outputting the calculated data frames and the rendered chart files; S40 performs domain knowledge retrieval and fault diagnosis. Based on the abnormal indicators or performance deviations calculated in S30, the agent orchestration and control module generates diagnostic question instructions. The knowledge base retrieval module recalls relevant domain rule texts according to the diagnostic question instructions. The domain knowledge diagnosis model receives the recalled text and the current abnormal indicators, performs fault association mapping, and outputs corresponding cause attribution and handling suggestions. S50, Perform factual consistency verification. The degradation fault-tolerant control module receives the summarized analysis conclusions and diagnostic text, and extracts all numerical characters contained in the text. The degradation fault-tolerant control module matches and verifies the extracted values ​​with the structured operational data obtained in S20, and adds a warning mark to values ​​that do not pass the tolerance matching standard, thereby confirming the factual basis of the final analysis report; S60, outputs a runtime diagnostic report or triggers a degradation response. Under normal execution conditions, the agent orchestration control module assembles and outputs the integrated data tables, chart files, and diagnostic conclusions after consistency verification according to a preset format. When the degradation and fault tolerance control module detects a timeout or cyclic call in any of the above steps, it triggers a backup degradation strategy, extracts the locally generated result data from the system cache pool, encapsulates it, and outputs a basic analysis report to terminate the current task flow.

[0032] See attached document Figure 2 In this embodiment, the specific execution process of receiving the run analysis request instruction and initializing the task context further includes the following steps: S101, in a specific implementation scenario, when the system starts, the intelligent agent orchestration and control module preloads system prompts. These prompts embed functional descriptions of each execution tool and formatted database schema information, thus constructing the basic operating environment. When an operator inputs a natural language request through the terminal interface, the system's human-machine interface receives unstructured text data containing inquiries about the operating status of thermal power plant equipment or anomaly diagnosis intentions. To avoid abnormal interruptions in subsequent model inference, the intelligent agent orchestration and control module performs format validation on the natural language request, removing illegal control characters contained in the text, thus completing the preprocessing operation of the basic text data.

[0033] In step S102, after acquiring the preprocessed text data, the intelligent agent orchestration and control module concatenates it with the aforementioned system prompts and inputs it into the encoding and decision-making reasoning model for intent parsing and entity extraction. Based on the operational analysis scenario of a thermal power plant, this model adopts an open-source large language model architecture pre-tuned with low-rank adaptive instructions. The specific dimensions of its input data are represented as discrete text word sequences. Internally, the model uses an attention mechanism to map and calculate the contextual relationships of unstructured text, transforming user intent into machine-recognizable structured business instructions. In the parsing phase, the encoding and decision-making reasoning model decomposes the key entity dimensions involved in the current task from the natural language request, specifically extracting equipment objects, time intervals, and analysis target indicators. This step establishes clear constraint boundaries for the subsequent generation of structured query language. If a key entity is found to be missing during parsing, the system will trigger an interactive warning, outputting a prompt request to guide the operator to supplement the relevant elements.

[0034] Taking the input text as an example of analyzing whether the nitrogen oxide emissions from the denitrification system of Unit 1 this month exceed the standard, the model decomposes and maps it into structured key-value pair information, extracts the equipment object as the denitrification system of Unit 1, the time interval as this month, and the analysis target indicator as the nitrogen oxide emission concentration. For the specific underlying network computation logic of entity recognition and word vector encoding mapping performed by the large language model, those skilled in the art can refer to the existing self-attention mechanism and feedforward neural network computation process of large language models, which are well-known technologies in the field and will not be elaborated here.

[0035] S103, based on the extracted structured entity information, the system needs to establish a persistent state tracking mechanism. The agent orchestration and control module allocates an independent session identifier for the current natural language request in the system memory pool, and encapsulates the extracted device object, time interval, and analysis target indicators into a structured context object, registering it as the initial global environment variable under this session identifier. This established context object remains resident throughout the lifecycle of the agent's subsequent multi-tool scheduling, serving as the baseline parameter for underlying data querying and code generation, and is called and updated across steps, thereby completing the underlying environment initialization configuration for the analysis task.

[0036] See attached document Figure 2 In this embodiment, the specific process of performing data query and quality assessment further includes the following steps: S201, In a preferred embodiment, after completing the task context initialization, the system needs to map the physical structure of the underlying industrial data source into text features that the model can understand through the database parsing component. The intelligent agent orchestration and control module calls the built-in database parsing component to scan the underlying industrial database, historical database, relational database, or time-series data view formed by mapping. The database parsing component automatically parses out the names, field comments, field data types, primary key identifiers, and their associated foreign key mapping relationships of all data tables in the database, and extracts no more than 5 unique example values ​​for each field to generate customized database schema description information.

[0037] Based on this, the agent orchestration and control module formats and assembles the task context object extracted in S10 with the parsed database pattern information to form query prompts for the model to read. This assembly process provides clear table structure boundary constraints for the conversion of intent into standard machine instructions.

[0038] S202, after obtaining the query suggestions, the system needs to convert them into a structured query language readable by the underlying database. The intelligent agent orchestration and control module inputs the assembled query suggestions into a dedicated text-to-SQL model. This model is built on an open-source large language model base, with its input features being discrete word sequences containing natural language instructions and database schema information, and its output being structured query code. To better adapt the general large model to the specific mapping task of the power plant's industrial control data table structure, the system pre-tunes the base model using a triplet dataset consisting of question text, structured query statements, and database schemas, performing low-rank adaptive instruction fine-tuning. During the fine-tuning training phase, the algorithm freezes the main parameters of the base model, injects low-rank matrices only into the relevant projection layers of the self-attention mechanism, and uses the cross-entropy loss function to calculate the distribution difference between the generated statements and the real structured query language labels. The weight parameters of the low-rank matrix are iteratively optimized using the gradient descent algorithm. The model outputs the corresponding structured query language statements based on the input information.

[0039] Subsequently, the intelligent agent orchestration and control module inputs the statement into an isolated execution sandbox configured with computational resource limitations. The isolated execution sandbox connects to the underlying database via a read-only database access interface and performs syntax type validation on the structured query language before execution, allowing only data query statements to pass; statements involving writes, deletions, updates, or structural changes are rejected. To balance data analysis performance with the security of the underlying core system, the isolated execution sandbox is instantiated as an independent container execution environment, with its physical memory limit pre-set to 512MB, computational resource quota limited to 50% of single-core processing power, and configured with execution timeouts, network access restrictions, and output directory restrictions. The system converts the query code to Base64 encoding format before decoding and execution in the isolated execution sandbox to reduce escaping anomalies caused by special characters during transmission or shell parsing. This operation retrieves the corresponding timing sensor parameters from the distributed control system or plant-level monitoring information system and returns them to the multi-dimensional running data acquisition module in the form of data frames. When processing multi-source heterogeneous data, the multi-dimensional running data acquisition module uses the time axis of high-frequency time series data as a reference and performs timestamp reference alignment for low-frequency offline feature parameters using the forward padding method to eliminate data misalignment caused by differences in sampling frequencies of different devices.

[0040] S203. Due to the complex physical environment of industrial sites, low-level sensors are prone to signal degradation and measurement distortion during long-term operation. Therefore, it is necessary to perform statistical anomaly detection on the retrieved data sequences. After acquiring structured operational data, the data cleaning unit built into the multidimensional operational data acquisition module performs mathematical and statistical analysis on the data sequences. To address common data degradation and distortion issues in industrial sites, the data cleaning unit establishes three feature identification standards.

[0041] The first is missing value identification, which involves traversing the dataset to find empty sites where no values ​​have been collected; the second is frozen value identification, which involves detecting data segments whose values ​​remain absolutely constant within a time window to identify constant value faults in sensor probes; and the third is abrupt value identification.

[0042] Based on the prior physical assumption that high-frequency sensor noise signals under steady-state operating conditions in thermal power plants approximately follow a normal distribution, the system uses the Laida criterion for judgment. The arithmetic mean and standard deviation of the sample data within the current time window are calculated. When the absolute deviation of a local data point from the arithmetic mean exceeds three times the standard deviation, the value is determined to deviate from the high-probability normal physical fluctuation range, and the system marks it as a sudden outlier caused by equipment flushing or electromagnetic interference. For the basic mathematical and statistical logic such as calculating the standard deviation, those skilled in the art can refer to existing general data analysis algorithms, which are well-known techniques in the field and will not be elaborated upon here.

[0043] S204, after completing the anomaly feature marking, the system determines the overall usability of the current data segment from a macro perspective. The multi-dimensional running data acquisition module counts the total number of the three types of abnormal data within the current time window, divides the total number of abnormal data by the total number of sample data within the current time window, and calculates the proportion of abnormal data. To prevent the denominator from approaching 0 during the calculation process and causing logical collapse, the system has a preset pre-interception mechanism. When the total number of sample data is detected to be equal to 0, the evaluation process is directly stopped and an error warning of an empty dataset is output.

[0044] When the calculated percentage of abnormal data exceeds a preset 10% threshold, the multi-dimensional operational data acquisition module determines that the reliability of this batch of data is too low, generates a data quality anomaly result, and writes the percentage of abnormal data, anomaly type, and involved fields into the current session context. The intelligent agent orchestration control module terminates the deep analysis calculation step based on this data quality anomaly result and calls the degradation fault tolerance control module to output a basic quality report. This basic quality report includes at least the query scope, sample size, percentage of abnormal data, distribution of anomaly types, and suggested review items. The basis for setting 10% as the interception threshold is that anomalies exceeding this limit in the industrial control data sequence of thermal power plants will cause subsequent thermodynamic statistical characteristics to shift, exceeding the repair range of conventional interpolation algorithms.

[0045] When the proportion of abnormal data does not exceed the 10% threshold, the system performs interpolation to repair missing parts of the dataset determined to be usable, and calls the moving average algorithm to smooth and denoise high-frequency fluctuation segments. After processing, the system performs a maximum-minimum mapping operation on the numerical results to complete the normalization transformation, thereby eliminating the physical scale differences in the dimension of different types of sensors. When the maximum and minimum values ​​of the field to be normalized are equal, the system skips the maximum-minimum mapping of that field and adds an unnormalized label to the field, or maps the field to a preset neutral value, outputting standardized structured operational data for subsequent steps.

[0046] See attached document Figure 3 In this embodiment, the specific process of performing depth analysis calculations and visualization rendering further includes the following steps: In step S301, after completing data querying and standardization, the system enters the mathematical statistics and visualization stage targeting core indicators. The intelligent agent orchestration and control module inputs the analysis target indicators parsed in S10 and the structured operational data output in S20 into the encoding and decision-making inference model. To ensure that the large language model possesses dedicated programming capabilities for industrial data analysis, the model is pre-built on an open-source large language model platform and undergoes low-rank adaptive instruction fine-tuning using a hybrid dataset containing natural language question text, statistical analysis code, chart rendering code, and inference decision samples. After receiving discrete text word sequence input, the model completes logical mapping in its feedforward neural network and outputs program code containing data processing logic and chart drawing instructions. This program code internally references statistical analysis code libraries such as pandas and numpy to perform data slicing, aggregation, and mathematical logic calculations, and calls chart rendering libraries to draw visualization charts such as trend charts or scatter plots reflecting changes in physical states.

[0047] In S302, to ensure the underlying system is not affected by potentially unknown code generated by large models, the system implements physical isolation and state management for code flow at the execution level. The agent orchestration and control module converts the generated program code into Base64 encoding format and passes it into the aforementioned isolated execution sandbox, which has pre-configured memory and computing resource limits, for decoding and execution. This operation reduces the risk of unescaped special characters causing parsing errors in the host operating system shell program. Simultaneously, when processing multi-step sequential analysis tasks, the system maintains variable state within the isolated execution sandbox through a serialization mechanism. The system calls a dedicated byte serialization interface to convert the core data frame variables calculated in the current step into a byte stream and write it to the resident buffer of the isolated execution sandbox. When the code in the next step is executed, it is read back and restored to a data frame structure through a deserialization mechanism. This cross-step variable sharing mechanism ensures that the context data is not broken during continuous analysis and calculation. Finally, the isolated execution sandbox outputs the processed data frames and rendered chart files.

[0048] S303. During multi-step inference computation driven by a large language model, due to the variability of complex operational logic, the agent may repeatedly call the same tool or get trapped in an invalid loop. To address this, the system introduces a multi-layered repeated call interception mechanism. The agent orchestration and control module monitors the agent's actions of calling various dedicated execution tools in real time. For each tool call, the system extracts the parameter lexicon set of the current intended action and simultaneously reads the parameter lexicon set of historical call records in the same session thread. As the first layer of defense, the system uses the MD5 hash algorithm to perform a basic comparison between the current parameters and historical parameters. If the calculated hash values ​​match, the current action is directly intercepted.

[0049] S304. Considering that large models are prone to generating redundant tool call instructions with different textual expressions but semantically equivalent meaning, a single hash comparison cannot effectively intercept implicit invalid duplicate calls. Therefore, as a second layer of defense, the agent orchestration control module uses Jaccard coefficients to calculate the semantic overlap between the current intent call parameters and historical call parameters. The mathematical expression for this calculation process is as follows: ; In the formula, This represents the calculated semantic similarity coefficient; This represents the set of parameter terms to be invoked in the current intent; This represents the set of parameter tokens representing historical call records within the same session thread; This indicates the number of lexical units contained in the intersection of two parameter lexical sets; This represents the number of lexical units contained in the union of two parameter lexical sets.

[0050] To ensure the stability of the algorithm, when the number of detected and contained lexical units is equal to 0, the system defaults to outputting a semantic similarity coefficient of 0, in order to avoid illegal calculation errors where the denominator approaches 0.

[0051] This formula quantifies the degree of semantic overlap between two action commands in terms of actual business requirements. After calculation, the agent orchestration and control module determines the semantic similarity coefficient. Whether the similarity threshold is greater than or equal to the system's preset threshold. This similarity threshold is set to 0.55, and the value is determined based on an empirical critical value obtained through testing with thermal power plant analysis command samples. This value can tolerate reasonable error correction iterations when the model explores the analysis path, and can effectively cut off invalid duplicate requests with high similarity.

[0052] If the second layer of defense is still bypassed, the system activates the third layer of defense, which verifies the cumulative number of calls made by a single dedicated execution tool in the current session thread. When the calculated semantic similarity coefficient is greater than or equal to the similarity interception threshold, or when the cumulative number of calls made by a single dedicated execution tool reaches the system's preset maximum call limit, the system marks the call as redundant and intercepts it. Once interception is triggered, the agent orchestration control module terminates the current invalid inference branch, and the instruction encoding and decision reasoning model directly outputs the interim analysis conclusions based on the existing execution context, thereby ensuring the overall operational analysis task continues to advance.

[0053] See attached document Figure 2In this embodiment, the specific process of performing domain knowledge retrieval and fault diagnosis further includes the following steps: In the knowledge base configuration phase of S401, to compensate for the lack of professional knowledge in vertical industrial scenarios using a general-purpose large language model, the system needs to pre-build a localized power plant-related knowledge base. The knowledge base retrieval module collects operating procedures, historical operating cases, and maintenance logs from thermal power plants, and divides the content into five major professional categories: noise and vibration issues, solid waste treatment issues, wastewater discharge issues, flue gas pollutant issues, and equipment and operating efficiency issues.

[0054] To achieve machine retrieval and computation of unstructured text, the knowledge base retrieval module calls a text segmentation component to slice and block the original text, and then calls an embedding model to map the discrete text blocks into high-dimensional dense vectors. The embedding model transforms the text word sequence into a continuous floating-point matrix representing contextual semantic features. Subsequently, the system attaches a corresponding professional category label as metadata to each high-dimensional vector and persistently stores it in the built-in vector database. This step aims to establish structured index partitions through metadata isolation, reducing the computational space for subsequent retrieval operations. For the specific network computation logic of the embedding model in performing the mapping from text words to high-dimensional vector space, those skilled in the art can refer to existing general natural language processing algorithms, which are well-known technologies in the field and will not be elaborated upon here.

[0055] In step S402, when the underlying equipment status indicators calculated in step S30 show abnormalities or performance deviations, the intelligent agent orchestration and control module generates a structured diagnostic query instruction based on the abnormal feature indicators. After obtaining this instruction, the system enters the hybrid similarity retrieval stage. Taking the specific operating condition of excessive nitrogen oxide emission concentration in the denitrification system as an example, the knowledge base retrieval module extracts the corresponding flue gas pollutant problem category tags based on the equipment and system attributes contained in the diagnostic query instruction and performs preliminary metadata filtering on the vector database. This filtering operation cuts off search paths for irrelevant categories.

[0056] After filtering, the knowledge base retrieval module also vectorizes the diagnostic query commands and uses an approximate nearest neighbor retrieval algorithm to obtain preliminary candidate texts within the selected vector space. During the retrieval calculation phase, the system uses Euclidean distance to measure the physical spatial distance between the query vector and the target vector in the database. The smaller the calculated distance value, the higher the semantic relevance between the two texts.

[0057] S403. Considering the large number of specialized terms and equipment codes in the thermal power industry that are difficult to replace with synonyms through simple semantics, semantic retrieval relying solely on Euclidean distance is prone to missing core technical terms. The knowledge base retrieval module further performs a re-ranking operation based on the BM25 keyword retrieval algorithm on the recalled preliminary candidate texts. The core of this algorithm lies in evaluating the weighted relationship between the density of specific words in a single document and their scarcity in the global corpus. It iterates through the key terms in the diagnostic query instructions, counts their frequency of occurrence in a single candidate text, and combines this with the inverse document frequency of the term in the entire knowledge base corpus to calculate a comprehensive matching weight coefficient.

[0058] To prevent statistical bias caused by extreme text lengths, the algorithm internally performs a normalization penalty on the length of the target text, avoiding unreasonably high scores for long and redundant documents. Based on this, the system fuses the semantic similarity spatial distance score calculated above with the keyword matching weight coefficient, rearranging the priority order of candidate texts, and thus outputting externally standardized text fragments to the subsequent model that balance contextual logical consistency and professional feature matching.

[0059] S404: After recalling external knowledge fragments, the system needs to perform the final logical mechanism attribution by combining the current real-time operating data. The intelligent agent orchestration and control module concatenates the recalled standard text with the current abnormal indicator data and inputs it into the domain knowledge diagnostic model within the dedicated model support module. To ensure that the model has professional reasoning capabilities for vertical industrial scenarios, the model is pre-tuned using low-rank adaptive instructions on the dataset using standard question-and-answer formats in the thermal power industry. During training, the algorithm freezes the main parameters of the open-source base model and only updates the weights by connecting a low-rank matrix in parallel at the self-attention mechanism layer. The cross-entropy loss function is used to calculate the difference between the generated text and the conclusions of real experts to achieve parameter optimization.

[0060] Its training corpus focuses on knowledge graphs of thermal power plant operation and maintenance, equipment structure knowledge such as rotor dynamics and thermodynamics principles, as well as the mapping relationship between fault types and characteristic parameters and expert experience knowledge. The model receives discrete word sequences containing prior factual background and real-time operating conditions as input, performs the association mapping between abnormal features and underlying physical mechanisms in its feedforward neural network, and finally outputs structured diagnostic conclusions. For the aforementioned scenario of excessive nitrogen oxide emissions, the specific business meaning output by the model not only includes the analysis of reasons such as low oxygen levels leading to incomplete combustion or decreased activity of denitrification catalysts, but also simultaneously outputs practical operation and maintenance troubleshooting suggestions for checking oxygen setpoints or detecting catalyst activity, thereby generating diagnostic conclusion text with industrial guidance value.

[0061] See attached document Figure 4 In this embodiment, the specific process of performing fact consistency verification further includes the following steps: S501, in the process of generating long text diagnostic conclusions using a large language model, due to the characteristics of its underlying autoregressive generation mechanism, there is a probability that the system will deviate from objective data, resulting in factual deviations. To ensure the rigor of the industrial analysis report, the system introduces a factual consistency verification mechanism based on underlying data tracing before outputting the final conclusion.

[0062] The degradation and fault tolerance control module invokes the built-in analysis conclusion verification tool, receives the analysis conclusions and diagnostic text summarized in the above steps, and uses preset regular expression matching logic to traverse the text, extracting quantitative numerical expressions referenced in the text, specifically including integers, decimals, and percentage values ​​with percentage signs. It also extracts surrounding text as context information. During extraction, the system labels the numerical type based on context keywords, unit information, and field aliases, and filters or separately identifies non-operational indicator values ​​such as drawing numbers, step numbers, equipment numbers, dates, and times, thereby converting the operational indicator values ​​in the diagnostic text into discrete numerical variables that can be mathematically verified. For the specific underlying code implementation of character matching in the regular expression matching logic, those skilled in the art can refer to existing general text parsing logic, which is well-known in the field and will not be elaborated here.

[0063] In step S502, after obtaining the extracted numerical variables and their context information from the text, the degradation and fault tolerance control module synchronously reads the original structured runtime data obtained in step S20 from the system cache, as well as the computational data frame generated in step S30 and the registered derived indicator results. The system reads all rows and columns of the structured query results through loop traversal logic, extracts the original numerical fields contained therein, and establishes a candidate matching set based on field name, field alias, unit information, and numerical context. Subsequently, the system performs a comparison operation, cross-matching the extracted text numerical variables with the original field values ​​or derived indicator values ​​in the candidate matching set.

[0064] Considering that large models typically generate reasonable floating-point truncation or rounding errors during the initial text generation or assembly code calculations, directly using mathematical equivalence comparison conditions could easily lead to false positives. Therefore, the degradation-tolerance control module sets an approximate matching tolerance standard or numerical precision matching rule with an absolute relative error of less than 1% in the comparison judgment logic. The 1% tolerance threshold is determined because this precision range can cover the mantissa deviation caused by normal floating-point arithmetic in computers, and is sufficient to eliminate abnormal values ​​that deviate from reality due to model calculation errors, ensuring the balance of fact-checking.

[0065] In S503, during the specific tolerance determination calculation process, the degradation tolerance control module calculates the absolute value of the difference between the text numerical variable and the original database field value, and divides it by the absolute value of the original database field value to obtain the relative error. To ensure the integrity of the algorithm logic and avoid illegal mathematical operations that cause the program to crash due to the denominator being equal to 0 during the calculation process, when the system detects that the original database field value used as the denominator is equal to 0, the evaluation algorithm automatically switches to the absolute error determination mode and determines whether the absolute difference between the text numerical variable and the original database field value is not greater than a preset absolute tolerance threshold. This preset absolute tolerance threshold is determined according to the minimum resolution of the corresponding sensor, the number of decimal places retained in the field, or the business configuration precision.

[0066] By comparing the data, for text values ​​that meet the criteria of relative error less than 1%, exact matching, or absolute error judgment, the system confirms that the feature value has real underlying physical data support, and the verification is passed. For abnormal values ​​that fail to match, the system records them in an exception queue. After the comparison is completed, the system automatically generates a structured verification report, which includes the number of matching numbers, the number of unmatched numbers, and a list of unmatched numbers. The system directly adds a warning label to the corresponding unmatched position in the diagnostic conclusion text to alert on-site maintenance personnel that the data lacks underlying basis.

[0067] In addition to directly referencing the underlying raw sensor data, the final diagnostic conclusion text of S504 typically includes derived indicators such as the proportion of total energy consumption and the performance ranking of similar devices, which require complex mathematical calculations to obtain. For these complex indicators that cannot be directly matched and verified with the underlying raw database, when simple numerical comparisons are insufficient to determine the factual basis, the degradation and fault tolerance control module will trigger a rollback mechanism, reactivating the system's internal coding and decision-making reasoning model for secondary logical cross-validation. The system extracts the generation results of the derived indicators and the basic structured data upon which the calculations depend, inputting them into the model. The instruction model then uses its own code reasoning capabilities to re-execute the reverse mathematical calculations.

[0068] If the final result obtained by reverse calculation is consistent with the value output in the diagnostic text, the complex indicator is deemed to have passed the verification; if there is a deviation in the calculation result, the system will also add a warning sign to the derived indicator, thereby ensuring the consistency of all quantitative descriptions in the analysis report output by the system.

[0069] See attached document Figure 5 In this embodiment, the specific execution process of outputting a runtime diagnostic report or triggering a degradation response further includes the following steps: In the S601, during multi-round interactions and analysis by the agent, the system faces the physical limitation of overflowing the large language model context window as historical records accumulate. To maintain the long-term semantic coherence of complex analysis tasks, a dynamic context memory hybrid compression algorithm is introduced into the degradation-tolerant control module. The system calls the built-in text encoding library to monitor the number of tokens in the current session history. The system uses a general text segmentation encoding scheme to perform the mapping and counting of tokens to digit sequences, and calculates the ratio of this mapping to the maximum capacity of the model context window in real time.

[0070] When the system detects that the ratio reaches or exceeds a preset threshold of 90%, it automatically activates the context compression module. The 90% threshold is set to ensure the system has sufficient space for the large language model's inference and feature mapping in the current round, avoiding unexpected physical truncation due to reaching physical limits. After compression is activated, the system employs a hybrid preservation strategy: it fully preserves the original interaction conversations from the most recent rounds while simultaneously using the encoding and decision-making inference models to generate a structured summary of no more than 300 characters from earlier historical conversations. This summary is explicitly required to include the main questions, key findings, and important decision information from the historical dialogues. The system then concatenates this structured summary with the preserved original conversations and re-injects it into the agent's context, thus mitigating the risk of semantic loss due to context truncation within limited physical memory constraints.

[0071] In addition to the hybrid strategy, the system also supports a truncation strategy that discards only earlier messages and a pure summary strategy, which can be configured by operations and maintenance personnel according to the working conditions.

[0072] In addition to limitations in physical computing resources, large language models face the risk of logical inference breaks when handling complex scenarios. To ensure the continuity of the entire process, the degradation and fault tolerance control module monitors the execution status of the agent in real time through a background monitoring component. The system performs semantic parsing on the last message returned by each large model communication, specifically monitoring three typical execution stall states: first, an empty response anomaly, where the model returns a result containing neither text content nor any tool call instructions; second, a transitional statement anomaly, where the model returns text containing only non-conclusive statements such as "the next step is analysis," "processing," or "continuing to query," without simultaneously generating executable tool call instructions or actionable interim conclusions; and third, unexpected termination of streaming execution.

[0073] S603: Once any of the above-mentioned abnormal warnings are triggered internally by the system, the degradation and fault tolerance control module immediately suspends the current normal inference process and initiates the state recovery mechanism. The system automatically extracts the core summary of the execution results of existing tools from past dialogue history and call trajectory logs, specifically including the row and column name information returned by structured queries, program code output values, local diagnostic conclusions, and saved file path information. The system reassembles the extracted running state summary and the initial task intent into new contextual prompts and re-injects them into the current execution thread to attempt to guide the agent to continue analysis and exploration based on existing results, avoiding a global system restart.

[0074] S604, if the system fails to recover after multiple consecutive attempts, the degradation and fault tolerance control module will determine that the current task flow cannot be closed under the normal large model inference path. The system presets a maximum number of recovery attempts of 3. When the number of failures reaches this limit, in order to ensure the robustness and high availability of the thermal power plant industrial analysis system, the system will activate the backup degradation strategy. The degradation and fault tolerance control module bypasses the autonomous planning layer of the agent and directly intervenes in the underlying system cache pool, automatically collecting the temporarily stored query statements and underlying return results, program code analysis output, domain knowledge diagnostic conclusions, and chart file lists, etc. Subsequently, the system assembles these results according to the preset standard markup language formatting template, directly encapsulates and outputs a basic analysis report to the specified isolated output directory, thereby ending the current task flow by outputting backup results.

[0075] S605, under normal execution conditions where the system does not trigger the aforementioned degradation strategy, when the agent successfully completes all planning steps and arrives at a complete conclusion, the agent orchestration control module will invoke the system's built-in file saving tool. This tool will assemble the integrated data tables, visualization charts, and the final diagnostic conclusions verified for factual consistency according to a standardized format, ultimately outputting a complete thermal power plant operation diagnostic report. The system will persistently store this operation diagnostic report in the designated business file system, allowing on-site maintenance personnel to access it as needed and perform corresponding physical equipment maintenance operations accordingly.

[0076] Specific application examples: Implementation background setting: This application example is set in the operating environment of Unit 1 of a 300MW thermal power plant. The system has pre-completed the vectorized storage of the knowledge base of equipment operation procedures and completed the schema information extraction from the relational database of the underlying distributed control system. The operation personnel observed abnormal fluctuations in environmental parameters on the monitoring interface and used the analysis system of this application to investigate the cause of the potential hazard.

[0077] Implementation steps: S10, the operator inputs a natural language request through the human-machine interface: Analyze the reasons for the increased nitrogen oxide emission concentration of the Unit 1 denitrification system during the period from 09:00 to 10:00 today. The intelligent agent orchestration and control module receives this request and calls the encoding and decision-making reasoning model for entity extraction. The model decomposes the intent into structured key-value pairs, extracts the equipment object as the Unit 1 denitrification system, the time interval as 09:00 to 10:00 (a total of 60 minutes), and the target indicator as nitrogen oxide emission concentration. Based on this, the system allocates an independent session identifier in the memory pool and initializes the task context.

[0078] S20, the dedicated text-to-SQL model generates query statements based on context and database schema and sends them to the isolated execution sandbox. It obtains discrete time-series data with a sampling frequency of 1 time / min from the underlying database, acquiring a total of 60 time nodes of nitrogen oxide concentration and boiler oxygen data.

[0079] The multidimensional operational data acquisition module performed statistical tests and identified two missing values ​​and one frozen value. In the abrupt change value identification stage, the system calculated the arithmetic mean of nitrogen oxide concentration during this period to be 45 mg / m³. 3 The standard deviation is 8 mg / m³. 3 According to the Raida criterion, the critical upper limit for normal fluctuations is 45 + 3 × 8 = 69 mg / m³. 3 The system detected a reading of 120 mg / m³ at the 15th minute. 3 The absolute deviation was 75, greater than 3 times the standard deviation, and the system marked it as a mutation outlier. A total of 4 outlier data points made the outlier percentage approximately 6.67% (4 ÷ 60 ≈ 6.67%). Since this percentage is less than the 10% interception threshold, the system determined the dataset was usable and used a moving average algorithm to smooth and correct the mutation value to the average of the preceding and following normal nodes, 48 ​​mg / m³. 3 This results in the output of standardized operational data.

[0080] S30, the agent orchestration and control module calls the model to generate program code for data statistics. During multi-step inference planning, the agent experiences action command redundancy, attempting to call the correlation calculation tool twice consecutively. The parameters recorded in the historical call log are for analyzing the correlation between nitrogen oxides and boiler oxygen levels, and their parameter term set... Extracted as {nitrogen oxides, boiler, oxygen content, correlation}; the current intended parameter is to analyze the correlation between boiler oxygen content and nitrogen oxides, and its parameter term set is... Extracted as {boiler, oxygen content, nitrogen oxides, association}. To intercept invalid duplicate calls, the system introduces the Jaccard coefficient to calculate its semantic overlap: ; In the formula, SJS_JSJ represents the calculated semantic similarity coefficient; This represents the set of parameter terms to be invoked in the current intent; This represents the set of parameter tokens representing historical call records within the same session thread; This indicates the number of lexical units contained in the intersection of two parameter lexical sets; This represents the number of lexical units contained in the union of two parameter lexical sets.

[0081] The system calculates that the intersection of the two sets contains 4 words, and the union of the two sets also contains 4 words. Substituting these values ​​into the formula yields the semantic similarity coefficient. =4÷4=1.0. Since 1.0 is greater than the preset interception threshold of 0.55, the system determines that the action constitutes a redundant call and intercepts it, automatically terminates the redundant branch, and directly outputs the calculated data frame (the Pearson correlation coefficient between the two is -0.85) and a trend chart file reflecting the negative correlation.

[0082] S40, targeting the aforementioned negative correlation features, the knowledge base retrieval module extracts the flue gas pollutant problem category label filtering vector database and uses an approximate nearest neighbor retrieval algorithm to obtain external normative text. The domain knowledge diagnostic model combines the recalled text and the calculation results, performs physical mechanism mapping in its feedforward neural network, and outputs a diagnostic conclusion: During this period, the boiler oxygen content decreased from 4.2% to 3.1%, leading to a deterioration of the combustion state in the furnace, the destruction of the reducing atmosphere, and an increase in the generation of nitrogen oxides.

[0083] In S50, the degradation and fault tolerance control module calls regular expression matching logic to extract numerical variables (such as 3.1) from the diagnostic text and traverses the raw structured runtime data in the system cache. The underlying database shows that the raw value of oxygen at the end of this period is 3.12%.

[0084] The system calculates the relative error between the two: the absolute difference (0.02) divided by the original value (3.12) yields a relative error of approximately 0.64%. Since this error is below the 1% approximate matching tolerance standard, the system confirms that the text feature value is supported by real underlying physical data and determines that the verification is successful.

[0085] S60: Since the current task flow has not triggered execution timeouts or loop stalls, the system does not need to activate the backup degradation strategy. The intelligent agent orchestration control module will assemble the integrated data tables, chart files, and final diagnostic conclusions that have undergone factual consistency verification according to the standardized format, and output a complete thermal power plant operation diagnostic report to the designated business file system.

[0086] Experimental verification and effect comparison: To verify the technical advantages of this application, a 30-day continuous operation dataset from the same batch of thermal power plants was used. This dataset contains 150 manually labeled real operational anomalies. The intelligent analysis method based on a large language model in this application was compared with the traditional system based on fixed threshold rules.

[0087] Comparison of anomaly attribution accuracy: Traditional methods rely solely on single threshold cross-alarms, lacking multi-parameter coupled calculations and external expert knowledge support, resulting in an attribution accuracy of 62.5% as tested. The method in this application, through data correlation calculations overlaid with a domain knowledge base for hybrid retrieval and diagnosis, achieves an actual attribution accuracy of 94.6%, effectively overcoming the one-sided judgment defect of single extreme values.

[0088] Comparison of false alarm rates for abnormal data: This includes data on sudden spikes caused by sensor rinsing or electromagnetic interference (such as 120 mg / m³ in S20). 3 Traditional methods directly trigger over-limit alarms, resulting in a false alarm rate of 18.2%. This application's method uses the Raida criterion for smooth noise reduction and interception, reducing the false alarm rate caused by sensor noise to 0%.

[0089] Fact verification validity: In the control group where the S50 fact consistency verification mechanism of this application was not enabled, the analysis report directly output by the large model showed a 7.3% fact deviation. After enabling the tolerance matching verification of this application, the system automatically intercepted and corrected false values ​​that deviated from the actual underlying data, and the fact consistency pass rate of the report reached 100%, ensuring the reliability of the analysis conclusions for industrial applications.

[0090] To visually verify the effectiveness of the proposed solution in abnormal data cleaning and false alarm rate control, a corresponding grayscale data trend verification chart is provided. Figure 6 .exist Figure 6 In the diagram, the horizontal axis represents the continuous sampling time of the equipment operation, measured in minutes; the vertical axis represents the nitrogen oxide concentration at the outlet of the denitrification system of the thermal power plant, measured in mg / m³. 3 . Figure 6 The main components include three key feature elements: the raw acquisition concentration (dashed line marked with a cross) represents the unprocessed raw feedback data from the underlying sensors, reflecting the abnormal spikes at specific time points (such as the 15th minute and the 45th minute); the smoothed and denoised concentration (dark solid line) represents the standardized operating data after being judged by the Laida criterion and processed by the moving average algorithm in this application, reflecting the true physical fluctuation trend after the system filters out abnormal spikes; and the preset control upper limit (flat dashed line) represents the upper limit of the control standard for power plant environmental emissions, used to compare whether the raw noise will cause the erroneous threshold alarm of traditional methods.

Claims

1. A method for analyzing the operation of thermal power plants based on AI large-scale models, characterized in that, Includes the following steps: The encoding and decision reasoning model is invoked to parse the intent of the received natural language request, extract the device object, time interval and analysis target indicators to complete the task context initialization; The system calls a dedicated text-to-SQL model and combines it with the database schema to generate a query statement. The query statement is then executed to obtain structured running data, and anomaly distribution checks are performed on the structured running data. For the structured runtime data that has passed the anomaly distribution test, the encoding and decision reasoning model is invoked to generate program code and execute it in an isolated execution sandbox, outputting data frames, chart files and anomaly indicators; Based on the abnormal indicators, a diagnostic question instruction is generated and the domain rule text is recalled. The domain rule text and the abnormal indicators are input into the domain knowledge diagnostic model to perform fault association mapping and output the diagnostic text. Numerical characters are extracted from the diagnostic text, and tolerance matching verification is performed between the numerical characters and the structured runtime data to generate a fact consistency verification result. Based on the factual consistency verification results, the data frame, the chart file, and the diagnostic text, a thermal power plant operation diagnostic report is output.

2. The method for analyzing the operation of thermal power plants based on a large AI model according to claim 1, characterized in that, The anomaly detection performed on the structured runtime data includes: Identify missing values, frozen values, and abrupt outliers in the structured operational data, and calculate the percentage of outliers in the total sample data. When the proportion of abnormal data exceeds a preset threshold of 10%, a basic quality report is output. When the proportion of abnormal data is not greater than a preset threshold of 10%, interpolation repair and smoothing denoising are performed on the structured running data, and maximum and minimum value mapping operation is performed to obtain the structured running data that passes the abnormal distribution test. The mutation outliers are determined by the Laida criterion, which calculates the absolute deviation of local data points from the arithmetic mean to be greater than 3 times the standard deviation.

3. The method for analyzing the operation of thermal power plants based on a large AI model according to claim 1, characterized in that, The process of calling the encoding and decision reasoning model to generate program code and executing it within an isolated execution sandbox also includes a mechanism to intercept repeated calls: Extract the parameter word set of the currently invoked action of the encoding and decision reasoning model, and the parameter word set of the historical invocation records in the same session thread; A hash algorithm is used to perform a basic comparison between the parameter word set of the current call to the encoding and decision reasoning model action and the parameter word set of the historical call records; When the comparison results are inconsistent, the number of words contained in the intersection of the parameter word set of the current call to the encoding and decision reasoning model action and the parameter word set of the historical call record, and the number of words contained in the union set are calculated. The number of words contained in the intersection set is divided by the number of words contained in the union set to obtain the semantic similarity coefficient. When the semantic similarity coefficient is greater than or equal to the preset similarity interception threshold of 0.55, or when the cumulative number of calls to the encoding and decision reasoning model reaches the preset maximum number of calls, the call action is intercepted and the current reasoning branch is terminated.

4. The method for analyzing the operation of thermal power plants based on an AI large model according to claim 1, characterized in that, The step of generating diagnostic query instructions and recalling domain rule text based on the abnormal indicators includes: The pre-acquired external specification documents are divided into different professional categories and sliced ​​into blocks. The discrete text blocks are mapped into high-dimensional dense vectors and attached with professional category labels for persistent storage in a vector database. Based on the anomaly indicators, a diagnostic question instruction containing a target category label is generated. The target category label in the diagnostic question instruction is extracted to perform metadata filtering on the vector database. An approximate nearest neighbor retrieval algorithm is used to obtain preliminary candidate texts and semantic similarity space distance scores based on Euclidean distance. The key terms in the diagnostic question instructions are traversed, the frequency of occurrence of the key terms in the preliminary candidate texts is counted, and the inverse document frequency in the entire knowledge base corpus is used to calculate the matching weight coefficient. The priority order of the preliminary candidate texts is rearranged by combining the semantic similarity spatial distance score and the matching weight coefficient, and the domain rule text is output.

5. The method for analyzing the operation of thermal power plants based on a large AI model according to claim 1, characterized in that, The step of extracting numerical characters from the diagnostic text, performing tolerance matching verification between the numerical characters and the structured runtime data, and generating a factual consistency verification result includes: The regular expression matching logic is called to extract the numerical characters in the diagnostic text, and the local text is used as context information to convert the numerical characters into discrete numerical variables. Iterate through the rows and columns of the structured running data to extract the original numerical fields and establish a candidate matching set; Calculate the absolute value of the difference between the discrete numerical variable and the original numerical field in the candidate matching set, and divide the absolute value of the difference by the absolute value of the original numerical field to obtain the relative error; When the relative error is less than 1%, the verification is deemed successful. When the relative error is not less than 1%, a warning mark is added to the corresponding unmatched position in the diagnostic text, and the factual consistency verification result is generated.

6. The method for analyzing the operation of thermal power plants based on a large AI model according to claim 1, characterized in that, The steps of extracting numerical characters from the diagnostic text, performing tolerance matching verification between the numerical characters and the structured runtime data, and generating a factual consistency verification result further include: Identify derived index values ​​containing the results of complex mathematical operations within the diagnostic text; Extract the generation results of the derived index values ​​and the basic structured data on which the calculation depends, and input them into the coding and decision reasoning model to perform reverse mathematical calculations; When the final result value obtained by the reverse mathematical calculation deviates from the derived indicator value output in the diagnostic text, a warning mark is added to the derived indicator value.

7. The method for analyzing the operation of thermal power plants based on an AI large model according to claim 1, characterized in that, After the task context initialization is completed, the following is also included: Monitor the number of tokens in the current session history corresponding to the task context, and calculate the ratio of the number of tokens to the maximum capacity of the context window of the encoding and decision reasoning model; When the ratio reaches or exceeds 90%, the context compression module is activated; The original interaction session of the most recent interaction round is retained, and the encoding and decision reasoning model is invoked to generate a structured summary of the earlier historical session; The structured summary is concatenated with the preserved original interaction session and then re-injected into the context of the encoding and decision reasoning model.

8. The method for analyzing the operation of thermal power plants based on a large AI model according to claim 1, characterized in that, The analysis method also includes an execution state recovery mechanism, including: The execution status of the encoding and decision reasoning model is monitored in real time by a background monitoring component; When an execution stall is detected, such as an empty response anomaly, a transitional statement anomaly, or an unexpected termination of streaming execution, the current regular inference process is paused, and the core summary of the execution results of existing external tools is extracted from the recorded current dialogue history and the large model call trajectory log. The core summary and the task intent obtained through intent parsing are assembled into a contextual cue word and then re-injected into the current execution thread.

9. The method for analyzing the operation of thermal power plants based on a large AI model according to claim 8, characterized in that, The analysis method also includes implementing a backup degradation strategy, including: When the number of failed attempts to re-inject the currently executing thread reaches the preset limit for recovery attempts, the system cache pool is accessed to collect temporary query statements, underlying return results, program code analysis output, domain knowledge diagnostic conclusions, and partial results of the chart file list. The collected local results are assembled according to a preset standard markup language template, and then packaged and output as a degraded power plant operation diagnostic report.

10. A thermal power plant operation analysis system based on an AI large-scale model, characterized in that, The method for analyzing the operation of thermal power plants based on an AI large model, as described in claim 1, includes: an intelligent agent orchestration and control module, used to perform the following operations: The encoding and decision reasoning model is invoked to parse the intent of the received natural language request, extract the device object, time interval and analysis target indicators to complete the task context initialization; The text-to-SQL conversion model is used in conjunction with the database schema to generate query statements. For structured operational data that passes the anomaly distribution test, the encoding and decision reasoning model is invoked to generate program code and executed in an isolated execution sandbox, outputting data frames, chart files, and anomaly indicators; Generate diagnostic query instructions based on the abnormal indicators; Based on the factual consistency verification results, the data frame, the chart file, and the diagnostic text, a thermal power plant operation diagnostic report is output. A multidimensional operational data acquisition module is used to run the query statement to obtain the structured operational data and perform anomaly distribution checks on the structured operational data; The knowledge base retrieval module is used to recall domain rule texts based on the diagnostic question instructions; A dedicated model support module is used to provide the encoding and decision reasoning model, the text-to-SQL dedicated model, and to input the domain rule text and the anomaly index into the domain knowledge diagnosis model to perform fault association mapping and output the diagnosis text; The degradation-tolerance control module is used to extract numerical characters from the diagnostic text, perform tolerance matching verification between the numerical characters and the structured running data, and generate the factual consistency verification result.

Images