Knowledge graph-based method, device and equipment for enhancing large language model for risk assessment of pressure-bearing equipment and medium

Abstract

The application relates to a pressure-bearing equipment risk assessment method and device based on a knowledge graph enhanced large language model, equipment and medium, the method comprising: forming a forced step-by-step deduction constraint based on a tree topology hierarchical solidified thinking chain reasoning path, enabling the large language model to perform hierarchical reasoning in a tree topology hierarchical order and output damage failure categories of each branch node and its subdivided damage modes in stages, obtaining damage failure probabilities of the pressure-bearing equipment according to a damage failure probability calculation model built in each leaf node; generating a risk assessment conclusion according to a failure possibility level corresponding to the damage failure probability and a failure consequence level corresponding to a damage failure constraint parameter; archiving the thinking chain reasoning path, industry standard clauses of each branch node, calculation bases of the damage failure probability and judgment bases of the risk assessment conclusion to form a traceable archive and output a natural description of the risk assessment of the pressure-bearing equipment.

Description

Technical Field

[0001] This application relates to the field of pressure equipment testing, and in particular to a method, apparatus, electronic device and computer-readable storage medium for risk assessment of pressure equipment based on a knowledge graph-enhanced large language model. Background Technology

[0002] Risk-Based Inspection (RBI), as a crucial technical means for the safety management of pressure equipment systems, has been widely applied in the integrity management of pressure equipment systems in petrochemical complete sets of plants. RBI optimizes the allocation of inspection resources from a "cycle-driven" to a "risk-driven" approach by systematically assessing the likelihood and consequences of failure in pressure equipment. Domestically, the GB / T 26610 series of standards has constructed a technical system covering the entire RBI process, from inspection strategy formulation and qualitative risk analysis to quantitative analysis of failure likelihood and consequences. Internationally, API RP 580 systematically standardizes the implementation process and management requirements of RBI, while API RP 581 further provides a quantitative assessment model for failure likelihood and consequences.

[0003] Currently, existing RBI methods still have the following technical shortcomings in engineering implementation, including: First, the risk assessment process relies excessively on human experience, resulting in insufficient consistency and traceability. Many crucial aspects of current RBI methods (such as corrosion loop delineation, damage mode identification, damage rate determination, and inspection strategy formulation) still heavily depend on the professional competence and experience of inspection personnel. The assignment of key parameters and risk assessment scales are easily influenced by subjective judgment, and different project teams may produce significantly different analysis results for the same pressure equipment system, severely hindering the efficiency of RBI implementation and the consistency of analysis results.

[0004] Secondly, domain knowledge is fragmented and difficult to reuse systematically. RBI involves a complex regulatory and standards system and fragmented cross-domain knowledge. Clauses, calculation processes, and engineering judgments in standards and specifications such as GB / T 26610 (Guidelines for Risk-Based Inspection of Pressure Equipment Systems), GB / T 30579 (Damage Pattern Recognition of Pressure Equipment), and API RP 571 (Damage Mechanisms of Fixed Equipment in the Refining Industry) are often scattered across different tools and documents in practice, lacking a unified information carrier. This makes it difficult to systematically reuse engineering experience and historical cases. Especially in the assessment of complex service conditions, atypical equipment, and equipment exceeding its design service life, there is a greater reliance on human experience for comprehensive judgment, which restricts the systematic implementation of RBI methods at the enterprise level.

[0005] Third, existing RBI software has limited knowledge integration and engineering interpretation capabilities. Commercial software, represented by DNV's Synergi RBI and mainstream domestic RBI software, integrates standard methodologies and risk models onto a unified platform to support the implementation and risk prioritization of enterprise-level RBI projects. However, such software primarily focuses on the implementation of methodological processes and calculation results, with relatively limited support for the systematic expression and reuse of the engineering logic behind standard clauses, the correlation of damage mechanisms, and historical project experience. Under complex service conditions, existing software still has room for improvement in terms of knowledge integration capabilities, process guidance, and engineering interpretation.

[0006] Fourth, general-purpose large language models are not sufficiently applicable in strictly regulated engineering scenarios. In recent years, existing large language model technologies have significantly improved their capabilities in natural language understanding, knowledge integration, and complex task processing, and have been used in fields such as medicine, finance, manufacturing, and energy for tasks such as document understanding, professional question answering, and decision support. However, pressure equipment RBI is a high-safety, strictly regulated scenario, and existing general-purpose large language models have shortcomings such as domain knowledge gaps, insufficient decision accuracy, "illusions," and poor interpretability, making it difficult to reliably function in key areas such as damage pattern recognition and inspection strategy analysis.

[0007] In summary, existing RBI methods rely heavily on human experience in engineering implementation, leading to insufficient evaluation consistency. Domain knowledge is scattered and difficult to reuse systematically. Commercial software knowledge integration and engineering interpretation capabilities are limited. Furthermore, general-purpose large language models suffer from knowledge gaps, illusions, and poor interpretability in strongly regulated scenarios. The applicant has made corresponding explorations to address these issues. Summary of the Invention

[0008] The purpose of this application is to solve the above problems by providing a method, device, electronic equipment and computer-readable storage medium for risk assessment of pressure equipment based on knowledge graph-enhanced large language model.

[0009] To achieve the various objectives of this application, the following technical solution is adopted: A risk assessment method for pressure equipment based on a knowledge graph-enhanced large language model, proposed to meet one of the purposes of this application, includes: The knowledge graph of pressure equipment is invoked, and each pressure equipment and its corresponding risk assessment data are used as matching conditions. The multi-level damage and failure classification structure of the knowledge graph is matched and filtered layer by layer to extract the subset of damage and failure classification nodes that are suitable for the pressure equipment and have a tree topology, and generate a multi-level damage and failure reasoning tree corresponding to the pressure equipment. The root node in the subset of damage and failure classification nodes is determined as a private root node, and its first intermediate branch node, second intermediate branch node, and leaf node are determined as shared nodes that can be reused by all pressure-bearing equipment. Damage and failure knowledge rules are uniformly bound within all shared nodes. Independent association links are established between the private root node and each shared node, and each association link stores damage and failure constraint parameters that are only suitable for the corresponding pressure-bearing equipment operating conditions in an isolated manner. Based on the tree-topology hierarchical fixed thinking chain reasoning path, a forced step-by-step deduction constraint is formed, which enables the large language model to perform hierarchical reasoning according to the tree-topology hierarchical order, outputting the damage failure category and its corresponding subdivided damage mode of each branch node level by level. After retrieving all the required parameters by using the damage mode output by the leaf node as an index, the damage failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage failure probability. Based on the failure probability level corresponding to the damage failure probability and the failure consequence level corresponding to the damage failure constraint parameter, the risk level of the target pressure equipment is determined, and a risk assessment conclusion is generated. The archived thought chain reasoning path, industry standard clauses corresponding to each branch node, calculation basis for damage failure probability, and judgment basis for risk assessment conclusion form a traceable archive, outputting a natural description of the risk assessment of pressure equipment.

[0010] Optionally, the damage failure classification node subset includes a private root node representing the pressure-bearing equipment, a first intermediate branch node representing the major damage failure category, a second intermediate branch node representing the subdivided damage mode, and a leaf node representing the damage failure probability. The pressure-bearing equipment includes a complete set of petrochemical plant pressure vessels and pressure pipelines, wherein the complete set of petrochemical plant pressure vessels includes a reaction vessel, a heat exchange vessel, a separation vessel and a storage vessel. The risk assessment data includes cumulative service time, number of actual detected damage types, degradation rate coefficient of each damage, operating condition fluctuation deviation coefficient, detection data missingness coefficient, material mechanical property parameters, and corrosion resistance parameter. The damage failure knowledge rules include corrosion mechanisms, industry standard clauses, damage failure probability calculation models, and detection and maintenance rules corresponding to each subdivided damage mode. The damage and failure constraint parameters include the operating condition boundary conditions, the range of medium composition limits, and the range of medium concentration limits.

[0011] Optionally, the step of performing layer-by-layer matching and filtering in the multi-level damage and failure classification structure of the knowledge graph to extract a subset of damage and failure classification nodes that are suitable for the pressure-bearing equipment and have a tree-like topology, and generating a multi-level damage and failure inference tree corresponding to the pressure-bearing equipment, includes: Feature extraction is performed on the pressure-bearing equipment and its corresponding risk assessment data to obtain a feature vector related to the risk assessment target of the pressure-bearing equipment. Based on the feature vector, locate the initial damage failure classification node that matches the risk assessment target in the knowledge graph of the pressure equipment; Starting from the initial damage failure classification node, following the multi-level damage failure classification structure in the pressure equipment knowledge graph, depth-first retrieval is performed level by level according to the preset thought chain reasoning path to extract the subset of damage failure classification nodes that constitute the multi-level damage failure reasoning tree; The subset of damage and failure classification nodes is organized into a tree topology according to hierarchical relationships to generate a multi-level damage and failure reasoning tree adapted to the pressure-bearing equipment.

[0012] Optionally, based on the fixed hierarchical thinking chain reasoning path of the tree topology, a mandatory step-by-step deduction constraint is formed, causing the large language model to perform hierarchical reasoning according to the tree topology hierarchy, and outputting the damage and failure categories of each branch node and their corresponding subdivided damage modes step by step, including: The structured entities, hierarchical relationships, node-embedded rules, and thought chain reasoning paths of the multi-level damage failure reasoning tree are input as constraint contexts into the large language model. The large language model traverses the damage and failure classification nodes of the multi-level damage and failure reasoning tree layer by layer from top to bottom along the thought chain reasoning path of the pressure-bearing equipment, damage and failure category, subdivided damage mode, and calculation of damage and failure probability in a step-by-step chain derivation. During the hierarchical derivation process, the damage failure knowledge rules and the damage failure constraint parameters are synchronously invoked to verify the reasoning process step by step. The reasoning branches generated during the reasoning process are verified according to industry standard clauses and working condition numerical boundary conditions to complete the topology constraint enhancement.

[0013] Optionally, after retrieving all required parameters using the damage mode output by the leaf node as an index, the damage failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage failure probability, including: Using the subdivided damage modes output by each leaf node as an index, the degradation rate coefficient and non-destructive testing reduction coefficient corresponding to each damage mode are retrieved in parallel in the shared node, and the damage co-coupling amplification index is obtained by combining and matching all identified damage types; the cumulative service time, operating condition fluctuation deviation coefficient and detection data missingness coefficient are retrieved from the private link; the baseline inherent failure probability and weight coefficient are retrieved from the shared node. Based on the number of actual detected damage types, the degradation rate coefficient, and the cumulative service time, calculate the multi-damage independent degradation cumulative term; The multi-damage coupling amplification factor is obtained by exponentiation of the damage synergistic coupling amplification index and the multi-damage independent degradation accumulation term; the hierarchical targeted inspection correction factor is obtained by multiplying the non-destructive testing reduction coefficient. The uncertainty correction factor is obtained by weighting and summing the working condition fluctuation deviation coefficient and its corresponding first weight coefficient, the detection data missingness coefficient and its corresponding second weight coefficient, and then adding the sum to the value. The damage failure probability of the pressure equipment is obtained by multiplying the baseline inherent failure probability, the multi-damage coupling amplification factor, the hierarchical targeted inspection correction factor, and the uncertainty correction factor.

[0014] Optionally, the step of determining the risk level of the target pressure-bearing equipment and generating a risk assessment conclusion based on the failure probability level corresponding to the damage failure probability and the failure consequence level corresponding to the damage failure constraint parameter includes: Obtain the working condition boundary conditions, medium component limitation range, and medium concentration limitation range from the damage failure constraint parameters; The failure consequence level of the target pressure-bearing equipment is determined based on the numerical boundary conditions of the operating conditions, the defined range of the medium components, and the defined range of the medium concentration. The failure probability is mapped to a preset failure probability level range to determine the failure probability level of the target pressure-bearing equipment; Based on the failure probability level and the failure consequence level, the risk coordinates of the target pressure-bearing equipment are located in a preset risk matrix. The risk matrix is ​​a two-dimensional risk assessment table constructed with the failure probability level as the vertical axis and the failure consequence level as the horizontal axis. Each intersection point in the two-dimensional risk assessment table corresponds to a risk level. The risk level corresponding to the risk coordinate is taken as the risk level of the target pressure-bearing equipment, and a risk assessment conclusion including the risk level determination result and the failure consequence level determination result is generated.

[0015] Optionally, the steps of archiving the thought chain reasoning path, the industry standard clauses corresponding to each branch node, the calculation basis for the probability of damage and failure, and the judgment basis for the risk assessment conclusion to form a traceable archive and output a natural description of the risk assessment of pressure equipment include: Collect the reasoning records generated at each level during the hierarchical reasoning process of the large language model along the tree topology, and form the thought chain reasoning path; Collect the original text and clause number of the industry standard clauses called by each branch node during the reasoning process, and form a standard clause call record; Collect the calculation basis for the damage failure probability and the judgment basis for the risk assessment conclusion, and form a conclusion support record; The reasoning path of the thought chain, the standard clause call record, and the conclusion support record are structurally associated according to the reasoning sequence to generate a traceable archive of the entire reasoning process; Based on the traceable archives, a natural language description of the risk assessment is output, including a summary of the reasoning chain, key judgment criteria, and risk assessment conclusions.

[0016] A pressure equipment risk assessment device based on a knowledge graph-enhanced large language model, provided for another purpose of this application, includes: The reasoning tree generation module is configured to call the knowledge graph of pressure equipment, use each pressure equipment and its corresponding risk assessment association data as matching conditions, match and filter layer by layer in the multi-level damage and failure classification structure of the knowledge graph, extract the subset of damage and failure classification nodes that are suitable for the pressure equipment and have a tree topology, and generate the multi-level damage and failure reasoning tree corresponding to the pressure equipment. The node configuration module is configured to define the root node in the subset of damage and failure classification nodes as a private root node, and define its first intermediate branch node, second intermediate branch node, and leaf node as shared nodes that can be reused by all pressure-bearing equipment. Damage and failure knowledge rules are uniformly bound within all shared nodes. Independent association links are established between the private root node and each shared node, and each association link stores damage and failure constraint parameters that are only suitable for the corresponding pressure-bearing equipment operating conditions in an isolated manner. The hierarchical reasoning module is set up to be a thought chain reasoning path based on a tree topology hierarchy, forming a mandatory step-by-step reasoning constraint. This allows the large language model to perform hierarchical reasoning in the order of the tree topology hierarchy, outputting the damage failure category and its corresponding sub-damage mode of each branch node level by level. After retrieving all the required parameters by using the damage mode output by the leaf node as an index, the damage failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage failure probability. The risk level determination module is configured to determine the risk level of the target pressure-bearing equipment based on the failure probability level corresponding to the damage failure probability and the failure consequence level corresponding to the damage failure constraint parameter, and generate a risk assessment conclusion. The traceable archive module is configured to archive the reasoning path of the thought chain, the industry standard clauses corresponding to each branch node, the calculation basis for the probability of damage and failure, and the judgment basis for the risk assessment conclusion, forming a traceable archive and outputting a natural description of the risk assessment of the pressure equipment.

[0017] An electronic device provided for another purpose of this application includes a central processing unit and a memory, the central processing unit being configured to invoke and run a computer program stored in the memory to perform the steps of the pressure equipment risk assessment method based on a knowledge graph-enhanced large language model described in this application.

[0018] A computer-readable storage medium is provided for another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the aforementioned risk assessment method for pressure equipment based on a knowledge graph-enhanced large language model, which, when invoked by a computer, executes the steps included in the corresponding method.

[0019] Compared to existing technologies, this application addresses the problems of existing RBI methods, such as insufficient evaluation consistency due to heavy reliance on human experience in engineering implementation, fragmented domain knowledge making systematic reuse difficult, limited knowledge integration and engineering interpretation capabilities of commercial software, and knowledge gaps, illusions, and poor interpretability of general-purpose large language models in strongly regulated scenarios. This application provides, but is not limited to, the following beneficial effects: Firstly, this application addresses the issue of excessive reliance on human experience in the risk assessment process, resulting in insufficient consistency and traceability. It establishes a rigid, tree-structured topology for the reasoning path, creating a mandatory step-by-step deduction constraint. This ensures the large language model strictly adheres to the hierarchical order of pressure-bearing equipment, damage / failure categories, subdivided damage modes, and leaf nodes, transforming the judgment process dependent on personal experience into a standardized, deterministic reasoning process. The reasoning conclusions of each branch node are based on industry standard clauses and knowledge rules uniformly bound within the shared node, avoiding assessment biases caused by differences in knowledge background and experience among different inspectors. Furthermore, by comprehensively archiving the reasoning path, standard clause citation records, and calculation / judgment basis through traceable documentation, the formation process of the assessment conclusions becomes verifiable and traceable, significantly improving the consistency and transparency of RBI analysis.

[0020] Secondly, this application addresses the issue of fragmented and unsystematic reuse of domain knowledge by constructing a knowledge graph for pressure equipment. This graph unifies the damage and failure knowledge rules from standards and specifications such as GB / T 26610 (Guidelines for Risk-Based Inspection of Pressure Equipment Systems), GB / T 30579 (Damage Pattern Recognition of Pressure Equipment), and API RP 571 (Damage Mechanisms of Fixed Equipment in the Refining Industry) into a globally shared node. Simultaneously, it structures and organizes historical RBI project reports, inspection records, and typical engineering cases to establish a case knowledge base, achieving systematic integration of fragmented knowledge. All pressure equipment shares the same set of knowledge rules within the shared node. New equipment only needs to configure its private constraint parameters to reuse the complete knowledge system, effectively reducing the marginal cost of knowledge maintenance and improving the reuse efficiency of engineering experience and historical cases.

[0021] Thirdly, addressing the limited knowledge integration and engineering interpretation capabilities of existing RBI software, this application uses a large language model as the inference engine. During the inference process, it synchronously calls damage and failure knowledge rules and constraint parameters for step-by-step verification, and performs compliance verification on the inference branches based on industry standard clauses and working condition boundary conditions, thus realizing a systematic expression of the engineering logic behind the standard clauses. By completely archiving the original text and number of the standard clauses called by each branch node through traceable archives, each inference conclusion has a clear engineering basis, significantly improving the engineering interpretation capability. At the same time, by establishing independent association links between private root nodes and each shared node, it realizes the separate management of general knowledge rules and individual constraint parameters, taking into account both the uniformity of the knowledge system and the flexibility of working condition adaptation, overcoming the limitations of existing commercial software that emphasizes process calculation and neglects knowledge integration and engineering interpretation.

[0022] Fourth, addressing the insufficient applicability of general-purpose large language models in strictly regulated engineering scenarios, this application uses a multi-level damage and failure reasoning tree as a structured retrieval enhancement index, thus limiting the knowledge sources and analytical scope of the large language model. It replaces the reasoning links dynamically generated by prompt words with a tree-topology-fixed hierarchical thought chain reasoning path, thereby suppressing model reasoning deviation and "illusion" risks from a mechanism perspective. Through a "dual-engine" architecture where the rule engine and the knowledge-enhanced large language model work together, knowledge rules and constraint parameters are synchronously invoked for step-by-step verification during the reasoning process. Furthermore, compliance checks are performed on the reasoning branches based on industry standard clauses and working condition boundary conditions. This forms a triple-credible reasoning mechanism: "domain knowledge graph constrains knowledge sources, fixed thought chain constrains reasoning paths, and rule engine constrains output results." This transforms the large language model from a potentially "illusionary" free text generator into a credible reasoning tool that strictly follows the hierarchical paths fixed by the engineering knowledge graph to perform forced step-by-step deductions, enabling the reliable application of the large language model in the high-safety, strictly regulated engineering scenario of pressure equipment RBI.

[0023] Furthermore, while maintaining the advantages of semantic understanding and natural language interaction of large language models, this application strictly constrains its reasoning behavior within the boundaries defined by engineering specifications and knowledge rules. This effectively suppresses the reasoning deviation and "illusion" risk of general large language models in strongly regulated engineering scenarios, and realizes the standardization of the evaluation process, the transparency of the reasoning path, and the traceability of the decision-making basis. It provides an intelligent risk assessment scheme for pressure equipment RBI that combines knowledge integrity, engineering constraints, and interpretability of results. Attached Figure Description

[0024] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1This is an exemplary network architecture used in the pressure equipment risk assessment system in this application embodiment; Figure 2 This is a flowchart illustrating the risk assessment method for pressure equipment based on a knowledge graph-enhanced large language model in this application embodiment; Figure 3 This is a flowchart illustrating the process of generating a multi-level damage and failure reasoning tree for a pressure-bearing device in an embodiment of this application. Figure 4 This is a schematic diagram illustrating the process of hierarchical reasoning of the large language model according to the tree topology in the embodiments of this application; Figure 5 This is a flowchart illustrating the process of calculating the damage failure probability of pressure-bearing equipment in an embodiment of this application. Figure 6 This is a schematic diagram of the process for generating risk assessment conclusions based on damage failure probability and damage failure constraint parameters in an embodiment of this application; Figure 7 This is a schematic diagram illustrating the natural description of the risk assessment of pressure-bearing equipment in the embodiments of this application; Figure 8 This is a schematic diagram of the pressure equipment risk assessment device based on a knowledge graph-enhanced large language model in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of the computer device in the embodiments of this application. Detailed Implementation

[0025] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0026] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0027] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0028] Those skilled in the art will understand that the terms "client," "terminal," and "terminal device" as used herein include both devices that receive wireless signals, devices that only possess wireless signal receiver capabilities without transmission capabilities, and devices with receiving and transmitting hardware, devices that have receiving and transmitting hardware capable of bidirectional communication over a bidirectional communication link. Such devices may include: cellular or other communication devices such as personal computers or tablets, having single-line displays, multi-line displays, or cellular or other communication devices without multi-line displays; PCS (Personal Communications Service) that can combine voice, data processing, fax, and / or data communication capabilities; PDAs (Personal Digital Assistants) that may include radio frequency receivers, pagers, internet / intranet access, web browsers, notebooks, calendars, and / or GPS (Global Positioning System) receivers; and conventional laptops and / or handheld computers or other devices that have and / or include radio frequency receivers. As used herein, "client," "terminal," and "terminal device" can be portable, transportable, installed in a means of transportation (air, sea, and / or land), or suitable and / or configured to operate locally and / or in a distributed manner, operating in any other location on Earth and / or in space. "Client," "terminal," and "terminal device" as used herein can also be a communication terminal, an internet access terminal, or a music / video playback terminal, such as a PDA, a MID (Mobile Internet Device), and / or a mobile phone with music / video playback capabilities, or a smart TV, set-top box, etc.

[0029] The hardware referred to by the names "server," "client," and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer. It is a hardware device with the necessary components revealed by the von Neumann architecture, such as a central processing unit (including an arithmetic logic unit and a control unit), memory, input devices, and output devices. The computer program is stored in its memory, and the central processing unit loads the program stored in the secondary storage into the main memory to run it, execute the instructions in the program, and interact with the input and output devices to complete specific functions.

[0030] It should be noted that the concept of "server" used in this application can also be extended to the case of server clusters. Based on the network deployment principles understood by those skilled in the art, the servers should be logically divided. Physically, these servers can be independent of each other but accessible through interfaces, or they can be integrated into a single physical computer or a computer cluster. Those skilled in the art should understand this flexibility and should not use it to constrain the implementation of the network deployment method in this application.

[0031] One or more of the technical features of this application, unless explicitly specified herein, can be deployed on a server and accessed by a client remotely calling the online service interface provided by the server, or can be directly deployed and run on a client for access.

[0032] Unless otherwise specified, the neural network models referenced or potentially referenced in this application may be deployed on a remote server and invoked remotely on the client, or deployed on a client with the capability to invoke directly. In some embodiments, when running on the client, the corresponding intelligence may be acquired through transfer learning in order to reduce the requirements on the client's hardware resources and avoid excessive consumption of the client's hardware resources.

[0033] Unless otherwise specified, all data involved in this application may be stored remotely on a server or on a local terminal device, as long as it is suitable for use by the technical solution of this application.

[0034] Those skilled in the art will understand that although the various methods in this application are described based on the same concept and thus present commonality among them, they can be performed independently unless otherwise specified. Similarly, the various embodiments disclosed in this application are all based on the same inventive concept; therefore, concepts expressed in the same way, as well as concepts that are appropriately changed for convenience but are expressed differently, should be understood equivalently.

[0035] Unless otherwise expressly stated, the various embodiments disclosed in this application can be combined in a cross-cutting manner to flexibly construct new embodiments, as long as such combination does not depart from the inventive spirit of this application and can meet the needs of the prior art or solve a certain deficiency in the prior art. Those skilled in the art should be aware of such modifications.

[0036] Please see Figure 1 The risk assessment method for pressure equipment based on a knowledge graph-enhanced large language model proposed in this application can be based on, for example... Figure 1 The pressure equipment risk assessment system shown is implemented. This system adopts a layered architecture, consisting of a hardware layer, a data layer, a software platform layer, a model and service layer, and an application layer, from bottom to top. Each layer interacts and collaborates with the others through standardized API / interface services, forming a complete closed loop from data collection, knowledge organization, intelligent analysis to business application and manual review. Specifically, the multi-source heterogeneous data provided by the data layer is cleaned and managed by the software platform layer. The knowledge management module then constructs a knowledge graph and damage / failure knowledge rule base for the pressure equipment RBI (Risk-Based Inspection) domain. After generating multi-level damage / failure reasoning trees for each pressure equipment based on the knowledge graph, the business process module uses them as a structured RAG (Retrieval Enhanced Generation) index, scheduling the large language model to execute forced step-by-step reasoning according to a fixed thought chain path; simultaneously, it calls the rule engine for constraint verification and compliance validation, forming a "dual-engine" collaborative working mechanism. The damage / failure probability and risk level output by the model and service layers are visualized through the results display module and simultaneously archived to form traceable files. After the inspectors complete the review and feedback at the application layer, the information is manually adjusted and sent back to the knowledge management module through the business process module. This information is used for the continuous supplementation of the knowledge base and the optimization and iteration of model configuration, forming a continuous improvement closed loop from intelligent analysis and manual confirmation to feedback optimization.

[0037] In some embodiments, the hardware layer includes computing resources, storage resources, network resources, and a deployment environment. The computing resources include general-purpose servers and GPU servers. General-purpose servers support the operation of business applications, rule engines, and API gateways; GPU servers provide high-performance parallel computing capabilities for inference and fine-tuning of large language models. The storage resources include database storage and file storage. Database storage stores structured data (e.g., equipment ledgers, inspection records, risk calculation results); file storage stores unstructured data (e.g., original text of standard clauses, historical inspection reports, case documents, etc.).

[0038] The network resources are deployed within the organization's internal network environment, enabling data exchange between different levels through intranet communication, while ensuring data transmission security and low latency. The deployment environment relies on supporting facilities such as a computer room and power supply to ensure stable 24 / 7 system operation, meeting the stringent availability requirements of special equipment inspection services.

[0039] In some embodiments, the data layer achieves unified collection and preprocessing of various types of data through a multi-source data access module. After unified storage and management, data query and sharing services are provided, offering a standardized data interface for upper-layer knowledge management and intelligent analysis. This data includes basic equipment data, operational and testing data, and domain knowledge data. The basic equipment data includes manufacturing information of pressure equipment (e.g., equipment type, design parameters, volume dimensions), material mechanical property parameters, and corrosion resistance parameters. This type of data originates from equipment design documents, material certificates, and factory documentation. The operational and testing data includes cumulative service time, historical operating condition records (e.g., temperature, pressure, media composition), alternating load characteristic spectrum, and records of all non-destructive testing and online monitoring. This type of data originates from equipment ledgers, DCS systems, and periodic inspection reports. The domain knowledge data includes relevant regulations and standards for RBI (Risk-Based Inspection) (such as TSG 21 (Safety Technical Inspection Code for Stationary Pressure Vessels), TSG 31 (Safety Technical Code for Industrial Pipelines), GB / T26610 (Implementation Guidelines for Risk-Based Inspection of Pressure Equipment Systems), GB / T 33578 (Detailed Rules for Risk-Based Inspection of Complete Equipment Sets), GB / T 30579 (Damage Pattern Recognition for Pressure Equipment), etc.), damage mechanism models, historical cases, and engineering experience. This type of data is the core source for constructing the domain knowledge graph and damage failure knowledge rule base.

[0040] In some embodiments, the software platform layer is the core business support platform of the system, including a data management module, a knowledge management module, a model service module, a business process module, a result display module, and a system management module. The data management module is responsible for cleaning, verifying, standardizing, and uniformly managing the multi-source heterogeneous data accessed by the data layer. Specific functions include: standardizing the format of basic equipment data and operational monitoring data to ensure consistency and availability of various types of data within the system; establishing a data quality control mechanism to identify and mark missing data and outliers; and providing standardized data access interfaces for the upper-layer knowledge management module and model service module.

[0041] The knowledge management module is responsible for the systematic organization and continuous updating of knowledge in the field of risk-based inspection (RBI) for pressure equipment. Specific functions include: constructing a computable and searchable knowledge graph for RBI of pressure equipment based on standard clauses, damage mechanisms, and case data provided by the data layer; establishing a dynamic update mechanism for the knowledge graph and standard specifications, performing structured analysis and version management of newly added or revised standard clauses; and structured organization and knowledge processing of historical RBI project reports, inspection records, and typical engineering cases to establish a case knowledge base.

[0042] The model service module is responsible for the invocation and management of the large language model and the rule engine. Specific functions include: encapsulating the API call interface of the large language model, providing semantic understanding, knowledge retrieval enhanced generation (RAG), and natural language generation services; managing the operation of the rule engine, executing deterministic calculation tasks such as damage trigger condition determination, key parameter range verification, and risk calculation models; and recording model call logs and inference processes to provide a data foundation for the generation of traceable archives. The business process module is responsible for the configuration and scheduling of the RBI (Risk-Based Inspection) risk assessment business process. Specific functions include: using the multi-level damage and failure reasoning tree generated from the knowledge graph as a structured RAG index, scheduling the large language model to execute hierarchical reasoning according to a fixed thought chain path; managing the human-machine collaborative workflow from "intelligent analysis to manual confirmation," controlling the start, pause, rollback, and completion of inference tasks; and realizing the workflow orchestration of the "dual-engine" collaborative scheduling of the rule engine and the large language model to ensure that the analysis process conforms to engineering specifications and safety boundaries. The results display module is responsible for the visualization and interaction of intelligent analysis results. Specific functions include: presenting the damage and failure categories, subdivided damage patterns, and risk levels output by the hierarchical reasoning of the large language model in a visual manner such as tree structures and risk matrices; displaying the original text of the standard clauses called during the reasoning process, the calculation basis, and the judgment basis, achieving an interpretable presentation of the analysis results; and providing a manual review interface to support inspectors in adjusting, supplementing, and confirming the intelligent recommendation results. The system management module is responsible for user management, access control, operation log recording, and system configuration management. Specific functions include: supporting multi-role user management (inspectors, reviewers, system administrators, etc.) and role-based access control; recording system operation logs to ensure the traceability of business operations.

[0043] In some embodiments, the model and service layer is the core of the system's intelligent decision-making, which includes rule and computation services, intelligent analysis services, and a knowledge base. The rule and computation services include damage mode and mechanism analysis, standard specifications and rule base, risk assessment and calculation, generation of inspection strategies and suggestions, constraint verification, and result sharing.

[0044] The intelligent analysis service is based on a large language model and achieves three core functions through a knowledge enhancement mechanism, including: (1) Semantic understanding and reasoning: Under the constraint of the thought chain reasoning path of the tree topology hierarchy, the reasoning is carried out step by step in the hierarchical order of "pressure-bearing equipment, damage failure category, subdivided damage mode, failure probability calculation", and the reasoning results of each branch node are output step by step to realize the constraint of forced step-by-step reasoning.

[0045] (2) Analysis and interpretation generation: Based on the structured retrieval index of the multi-level damage failure reasoning tree, standard clauses, cases and engineering experience in the knowledge base are called to generate explanatory texts on the formation mechanism of damage mode, risk influencing factors and reasoning conclusions.

[0046] (3) Knowledge retrieval: By enhancing the generation mechanism through retrieval, standard clauses, cases and engineering knowledge related to the current working conditions and damage modes are retrieved in real time in the knowledge base during the reasoning process, so as to ensure that the output of the large language model is based on evidence and traceable.

[0047] The knowledge base is a knowledge resource pool for the model and service layers, including the following three categories: (1) The standard clause library includes structured clause texts, judgment rules and threshold conditions of domestic and foreign standards such as GB / T 26610 (Guidelines for Risk-Based Inspection of Pressure Equipment Systems), GB / T30579 (Damage Pattern Recognition of Pressure Equipment), API RP 571 (Damage Mechanism of Fixed Equipment in Refining Industry), API RP 580 / 581, etc. Among them, API RP 580 / 581 is the core standard of risk-based inspection (RBI) issued by the American Petroleum Institute (API) and widely adopted worldwide.

[0048] (2) The engineering experience database includes historical RBI (risk-based inspection) project reports, typical failure cases, inspection records and expert experience, which are structured and used for case retrieval and experience reference.

[0049] (3) The rule knowledge base includes general knowledge rules given by industry standards, such as corrosion mechanism, damage synergistic coupling amplification index, non-destructive testing reduction coefficient and uncertainty weight coefficient.

[0050] In some embodiments, the application layer is the system's direct interactive interface for inspection personnel and engineers, providing full-process business function support for RBI (Risk-Based Inspection) risk assessment. It includes equipment information management, risk assessment and analysis, inspection strategy recommendation, analysis result display and interpretation, report generation and export, manual review and feedback, and a user interaction layer. Among them, the equipment information management supports the entry, query, maintenance, and management of pressure equipment ledgers, covering basic equipment information (type, material, design parameters), service information (date of commissioning, cumulative running time), inspection history, and maintenance and modification records, providing basic data input for risk assessment.

[0051] The risk assessment and analysis guides inspection personnel to complete risk assessments according to the RBI (Risk-Based Inspection) standard process. After manual confirmation of equipment information and operating parameters, the system automatically calls the model and service layer to complete knowledge graph matching, reasoning tree generation, hierarchical reasoning of the large language model, failure probability calculation, and risk level determination. The system also automatically archives the thought chain reasoning path, standard clause call records and calculations, and judgment basis during the analysis process.

[0052] The recommended inspection strategy is automatically generated based on the risk assessment conclusion and the inspection rule base. It includes: recommended inspection methods (such as ultrasonic testing, X-ray testing, acoustic emission, etc.), inspection cycle suggestions, key inspection areas and inspection coverage requirements, providing decision-making reference for inspection personnel to formulate scientific inspection plans.

[0053] The analysis results presentation and interpretation will visualize the risk assessment process and results, including: a multi-level damage failure reasoning tree, damage pattern recognition results and judgment criteria for each branch node, failure probability calculation process and parameter value explanation, risk matrix coordinates and risk level judgment results, and original texts of industry standard clauses and engineering case references called in the reasoning process.

[0054] The report generation and export function supports generating a standard-format risk assessment report from the entire risk assessment process data with one click, including: a summary of basic equipment information, damage pattern identification results, failure probability and risk level calculation results, inspection strategy recommendations, and a complete traceable file (including reasoning path, standard clauses, and calculation basis).

[0055] The manual review and feedback system provides a human-machine collaborative decision-making interface, supporting inspectors in reviewing, adjusting, and confirming the results of intelligent analysis. Specific functions include: manual correction of damage pattern recognition results, verification and confirmation of risk level assessment results, adjustment and supplementation of intelligently recommended testing strategies, and recording and feedback of manual review opinions. The content of manual adjustments and confirmations will continuously optimize the knowledge base and model configuration through a feedback mechanism. The user interaction layer indicates that the system is deployed in a web-based manner, providing a unified user portal for inspection personnel and engineers. Users access the system through a browser to complete the entire process of equipment information entry, risk assessment task initiation and tracking, analysis result viewing and review, report generation and export, etc., realizing the digital and intelligent management of RBI (Risk-Based Inspection) risk assessment.

[0056] Based on the above exemplary scenario, please refer to Figure 2 In one embodiment of the knowledge graph-based enhanced large language model risk assessment method for pressure equipment of this application, the method includes: Step S10: Call the knowledge graph of pressure equipment, and use each pressure equipment and its corresponding risk assessment data as matching conditions. In the multi-level damage and failure classification structure of the knowledge graph, match and filter layer by layer to extract the subset of damage and failure classification nodes that are suitable for the pressure equipment and have a tree topology, and generate the multi-level damage and failure reasoning tree corresponding to the pressure equipment. The pressure equipment risk assessment system in the terminal equipment can respond to pressure equipment risk assessment events, call the pressure equipment knowledge graph, and use each pressure equipment and its corresponding risk assessment associated data as matching conditions. It performs layer-by-layer matching and filtering in the multi-level damage and failure classification structure of the knowledge graph, extracting a subset of damage and failure classification nodes that are suitable for the pressure equipment and have a tree-like topology. It then generates a multi-level damage and failure reasoning tree corresponding to the pressure equipment. The subset of damage and failure classification nodes includes a private root node representing the pressure equipment, a first intermediate branch node representing the major damage and failure category, a second intermediate branch node representing the subdivided damage mode, and leaf nodes representing the damage and failure probability. The pressure equipment includes pressure vessels and pressure pipelines in a complete petrochemical plant. The pressure vessels in the complete petrochemical plant include reaction vessels, heat exchange vessels, separation vessels, and storage vessels. The risk assessment associated data includes cumulative service time, the number of actual detected damage types, the degradation rate coefficient of each damage, the operating condition fluctuation deviation coefficient, the detection data missingness coefficient, material mechanical property parameters, and corrosion resistance characteristic parameters. In some embodiments, the cumulative service time represents the cumulative operating time of the pressure equipment from commissioning to the time of risk assessment, used to quantify the total degradation of various types of damage over time. The number of actually detected damage types represents the total number of damage mode types currently existing on the equipment, confirmed based on the most recent valid in-service inspection, used to determine the traversal range of damage coupling effects in the failure probability calculation. The degradation rate coefficient of each damage represents a quantitative indicator describing the rate of deterioration of a specific damage mode over time (e.g., annual corrosion rate of uniform corrosion, fatigue crack propagation rate, etc.), used to calculate the cumulative degradation of the damage within the service cycle. The operating condition fluctuation deviation coefficient represents a quantitative indicator reflecting the degree to which key process parameters such as actual operating temperature and pressure deviate from the design operating conditions, used to correct for additional failure risks caused by fluctuations in operating conditions. The detection data missingness coefficient represents a quantitative indicator characterizing the completeness of historical inspection records of the equipment (e.g., wall thickness, crack, corrosion data), used to correct for assessment uncertainties introduced by incomplete data information. Material mechanical property parameters represent key indicators describing the mechanical behavior of materials used in pressure equipment (such as tensile strength, yield strength, impact toughness, etc.), used to determine the suitability of materials and their susceptibility to various types of damage. Corrosion resistance parameters describe the ability of materials used in pressure equipment to resist corrosion degradation under specific media and environmental conditions (such as corrosion-resistant alloy content, stress corrosion cracking susceptibility, etc.), used to determine the compatibility of materials with the service medium and operating conditions.

[0057] In some embodiments, the damage failure category refers to a primary classification of various failure modes that may occur in pressure-bearing equipment according to the common characteristics of damage mechanisms and failure forms, which is a coarse-grained classification of failure types. In the multi-level damage failure reasoning tree, the damage failure category serves as the first intermediate branch node, used to broadly classify the failure paths that pressure-bearing equipment may be involved in, and to determine which types of risk paths exist for the equipment. The damage failure category includes, but is not limited to, corrosion thinning, environmental cracking, material degradation, mechanical damage, and high-temperature damage, etc. Each damage failure category is further subdivided into various specific damage modes.

[0058] The subdivided damage pattern refers to an indivisible terminal damage type further subdivided within the framework of the major damage failure category. It represents the finest granular level in the damage failure classification system. In the multi-level damage failure reasoning tree, the subdivided damage pattern serves as a second intermediate branch node, used to accurately locate the specific damage type currently existing in the equipment after the major failure category is determined. The detailed damage modes described directly correspond to the damage mode definitions in industry standards (such as GB / T30579). Taking "corrosion thinning" as an example, its sub-damage modes include uniform corrosion, pitting corrosion, high-temperature oxidation, and erosion corrosion. Taking material degradation as a major category of damage failure, its sub-damage modes include graphitization, spheroidization, sensitization-intergranular corrosion, σ-phase embrittlement, high-temperature hydrogen damage, and tempering embrittlement. Graphitization refers to the decomposition of cementite into free graphite in carbon steel during high-temperature service. Spheroidization refers to the spheroidization of pearlite lamellar structure leading to a decrease in strength. Sensitization-intergranular corrosion refers to corrosion caused by chromium depletion at grain boundaries after sensitization of stainless steel. σ-phase embrittlement refers to the precipitation of σ-phase in stainless steel at high temperatures, leading to embrittlement. High-temperature hydrogen damage refers to the reaction of hydrogen with carbides in a hydrogen-rich environment, leading to decarburization and cracking. Tempering embrittlement refers to the segregation of impurity elements at grain boundaries in chromium-molybdenum steel, leading to a decrease in toughness.

[0059] In some embodiments, please refer to Figure 3 The steps of matching and filtering layer by layer in the multi-level damage and failure classification structure of the knowledge graph, extracting a subset of damage and failure classification nodes that are suitable for the pressure-bearing equipment and have a tree-like topology, and generating a multi-level damage and failure inference tree corresponding to the pressure-bearing equipment include: Step S101: Extract features from the pressure-bearing equipment and its corresponding risk assessment data to obtain feature vectors related to the risk assessment objectives of the pressure-bearing equipment; Since the original data, such as equipment type, material, cumulative service life, number of damage types, and degradation rate coefficient, are unstructured or semi-structured engineering data, they cannot be directly matched and located in the knowledge graph. This step transforms them into numerical feature vectors of a unified dimension through feature extraction, establishing a computable mathematical mapping relationship between equipment information and knowledge graph nodes. Each dimension of the feature vector corresponds to different attributes of the equipment (such as material category, temperature range, media type, etc.), providing a numerical matching basis for the next step of locating the starting node in the knowledge graph.

[0060] Step S102: Based on the feature vector, locate the initial damage failure classification node that matches the risk assessment target in the knowledge graph of the pressure equipment; The similarity between the feature vector of the equipment and the attribute vector of each node in the knowledge graph is calculated, and the node with the highest similarity is selected as the starting node. The initial damage failure classification node is the entry point for the generation of the reasoning tree, determining where subsequent reasoning will begin to unfold downwards in the knowledge graph. Different types of pressure equipment have different starting nodes (e.g., hydrogenation reactors enter from the "hydrogenation reactor damage mode" node, pressure pipelines enter from the "pipeline damage mode" node, etc.), achieving adaptive matching for different equipment types.

[0061] Step S103: Starting from the initial damage failure classification node, along the multi-level damage failure classification structure in the pressure equipment knowledge graph, perform depth-first retrieval step by step according to the preset thinking chain reasoning path to extract the subset of damage failure classification nodes that constitute the multi-level damage failure reasoning tree. Starting from the initial node located in step S102, the search proceeds level by level downwards according to the thought chain: pressure equipment, failure category, subdivided damage mode, and failure probability calculation. The depth-first strategy involves starting from a failure category node, searching along that branch to the lowest level of subdivided damage modes and leaf nodes, extracting complete single failure paths, and then backtracking to the next branch, sequentially traversing all candidate failure categories to ensure the reasoning tree covers all possible failure paths without missing any branches. At each node visited, the node and its attribute information are extracted, accumulating level by level to form a node subset.

[0062] Step S104: Organize the subset of damage and failure classification nodes into a tree topology according to hierarchical relationships to generate a multi-level damage and failure reasoning tree adapted to the pressure-bearing equipment.

[0063] Since the node subset extracted in step S103 is a flat list, it is merely an unordered collection of nodes and does not contain information on the hierarchy and subordinate relationships between nodes. This step reconstructs the flat list according to the inherent "parent-child" hierarchical relationship in the knowledge graph, forming a tree-like topology with the private root node as the root, the damage / failure category as the first-level branch, the subdivided damage pattern as the second-level branch, and the leaf node as the terminal, generating a structured reasoning index that can be directly called by the large language model. The tree-like topological hierarchy of the reasoning tree solidifies the thought chain reasoning path, providing the large language model with a structured knowledge retrieval index and a path constraint template for forced step-by-step reasoning.

[0064] As can be seen from steps S101 to S104 above, a deterministic transformation from "unstructured engineering data input" to "structured tree-shaped reasoning index output" has been achieved. This provides a structured data foundation with both knowledge integrity and engineering constraints for subsequent downstream tasks such as forced hierarchical reasoning of the large language model, leaf node calling of the damage failure probability calculation model, risk level determination, and traceable archive generation. This realizes the transformation of the large language model from a free text generator that may produce "illusions" to a reliable reasoning tool that strictly follows a fixed path to perform inference.

[0065] In some embodiments, the pressure-bearing equipment includes a complete set of petrochemical plant pressure vessels and pressure pipelines, wherein the complete set of petrochemical plant pressure vessels includes reaction vessels, heat exchange vessels, separation vessels, and storage vessels; the risk assessment associated data includes cumulative service time, the number of actual detected damage types, the degradation rate coefficient of each damage, the operating condition fluctuation deviation coefficient, the detection data missingness coefficient, material mechanical property parameters, and corrosion resistance characteristic parameters; the damage failure knowledge rules include corrosion mechanisms, industry standard clauses, damage failure probability calculation models, and detection rules and maintenance rules corresponding to each subdivided damage mode; the damage failure constraint parameters include operating condition numerical boundary conditions, medium component limitation ranges, and medium concentration limitation ranges.

[0066] Step S20: Determine the root node in the subset of damage and failure classification nodes as a private root node, and determine its first intermediate branch node, second intermediate branch node and leaf node as shared nodes that can be reused by all pressure-bearing equipment. All shared nodes are uniformly bound with damage and failure knowledge rules. Establish independent association links between the private root node and each shared node. Each association link stores damage and failure constraint parameters that are only suitable for the corresponding pressure-bearing equipment operating conditions in an isolated manner. The knowledge graph of pressure equipment is invoked, and each pressure equipment and its corresponding risk assessment data are used as matching conditions. The knowledge graph's multi-level damage and failure classification structure is matched and filtered layer by layer to extract a subset of damage and failure classification nodes that are suitable for the pressure equipment and have a tree-like topology. After generating a multi-level damage and failure inference tree for the pressure equipment, the root node in the subset of damage and failure classification nodes is determined as a private root node, and its first intermediate branch node, second intermediate branch node, and leaf node are determined as shared nodes that can be reused by all pressure equipment. Damage and failure knowledge rules are uniformly bound within all shared nodes. Independent association links are established between the private root node and each shared node. Each association link stores damage and failure constraint parameters that are only suitable for the corresponding pressure equipment's operating conditions in an isolated manner. The damage and failure knowledge rules include corrosion mechanisms, industry standard clauses, damage and failure probability calculation models, and detection rules and maintenance rules corresponding to each subdivided damage mode. The damage and failure constraint parameters include operating condition boundary conditions, media component limits, and media concentration limits.

[0067] Specifically, the first intermediate branch node, the second intermediate branch node, and the leaf node in the inference tree are designated as shared nodes that can be reused by all pressure-bearing equipment. This means that all equipment shares the same failure category classification system, the same subdivided damage mode library, and the same damage failure probability calculation model, ensuring the uniformity and consistency of the knowledge system throughout the entire system. All shared nodes are internally bound with uniform damage failure knowledge rules, including corrosion mechanisms, industry standard clauses, damage failure probability calculation models, and detection and maintenance rules corresponding to each subdivided damage mode. These rules are derived from industry standards and professional literature, representing universal knowledge applicable to all pressure-bearing equipment, and do not change with specific equipment operating conditions, ensuring the standardization and compliance of knowledge retrieval.

[0068] Independent links are established between the private root node and each shared node. Each link is bound to damage and failure constraint parameters that are only applicable to the specific operating conditions of the pressure equipment, including numerical boundary conditions of the operating conditions, and the limiting range of medium composition and concentration, using isolated storage. The links are isolated from each other, and the private parameters of one device will not affect the calculation logic of another device.

[0069] Because each pressure vessel operates under different conditions and media, its corresponding constraint parameters also differ. Through an isolated storage mechanism, each vessel's private parameters only take effect within its own inference path, preventing cross-device interference and balancing the uniformity of the knowledge system with the flexibility of engineering adaptation.

[0070] The ingenuity of the above architecture lies in separating the unchanging "rules" from the variable "parameters". Shared nodes ensure the standardization and compliance of knowledge invocation, while private association links ensure the flexibility and security of parameter adaptation. This allows the large language model to call the general rule base during inference, while also accurately incorporating the constraints unique to the current device.

[0071] In some specific embodiments, the corrosion mechanism is used as the fundamental theory to describe the physical / chemical failure processes and evolution laws of various damage modes under different materials, media, and operating conditions. It is used to determine whether a damage mode is likely to be triggered and its evolution path under specific operating conditions. The industry standard clauses are derived from specific judgment criteria and threshold provisions in regulations and standards such as TSG 21 (Safety Technical Supervision Regulations for Stationary Pressure Vessels), GB / T 26610 (Guidelines for Risk-Based Inspection of Pressure Equipment Systems), GB / T 30579 (Damage Mode Recognition for Pressure Equipment), and API RP 581, serving as authoritative bases for damage mode identification, risk quantification calculation, and compliance determination.

[0072] The damage failure probability calculation model is used to map equipment operating parameters and damage modes into quantified failure probability values. The detection rules specify recommended non-destructive testing methods (such as ultrasonic, radiographic, and acoustic emission) for various sub-type damage modes, along with detection coverage requirements and detection cycle determination rules, to guide inspection personnel in developing targeted inspection plans. The repair and handling rules, when the degree of a certain sub-type damage exceeds the standard allowable value, provide corresponding repair method selection rules (such as grinding removal, weld repair, partial replacement, etc.) and post-repair re-inspection and re-evaluation requirements, to guide maintenance decisions.

[0073] In some specific embodiments, the operating condition boundary conditions represent the limit values ​​and rated ranges of key process parameters such as temperature, pressure, and flow rate during the actual operation of the pressure equipment. These are used to determine whether the current operating condition of the equipment is within the design allowable boundaries; if the boundaries are exceeded, an abnormal inference branch interception is triggered. The medium component limitation range represents the type and content range of chemical components (such as hydrogen, hydrogen sulfide, carbon dioxide, chloride ions, etc.) contained in the internal medium of the equipment. This is used to determine whether a specific corrosive component has reached the concentration threshold condition for triggering a certain type of damage mode. The medium concentration limitation range represents the specific concentration range and fluctuation range of various corrosive components in the medium, and, in conjunction with the medium components, is used for the precise screening of damage modes and the quantitative determination of degradation rate coefficients.

[0074] Step S30: Based on the tree-topology hierarchical fixed thinking chain reasoning path, a forced step-by-step reasoning constraint is formed, so that the large language model performs hierarchical reasoning according to the tree-topology hierarchical order, and outputs the damage failure category and its corresponding subdivided damage mode of each branch node level by level. After retrieving all the required parameters by using the damage mode output by the leaf node as the index, the damage failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage failure probability. The root node in the subset of damage and failure classification nodes is designated as a private root node, and its first intermediate branch node, second intermediate branch node, and leaf node are designated as shared nodes that can be reused by all pressure-bearing equipment. Damage and failure knowledge rules are uniformly bound within all shared nodes. Independent association links are established between the private root node and each shared node. After each association link stores damage and failure constraint parameters that are only suitable for the corresponding pressure-bearing equipment operating conditions in isolation, a forced step-by-step deduction constraint is formed based on the thought chain reasoning path of the tree topology hierarchy. This allows the large language model to perform hierarchical reasoning according to the tree topology hierarchy, outputting the major damage and failure categories and their corresponding subdivided damage modes of each branch node level by level. After retrieving all the required parameters using the damage modes output by the leaf nodes as indexes, the damage and failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage and failure probability.

[0075] In some embodiments, please refer to Figure 4 Based on the fixed hierarchical tree-like topology reasoning path, a mandatory step-by-step deduction constraint is formed, enabling the large language model to perform hierarchical reasoning according to the tree-like topology, and outputting the damage and failure categories and their corresponding sub-damage modes of each branch node level by level, including: Step S301: Input the structured entities, hierarchical relationships, node built-in rules, and thought chain reasoning paths of the multi-level damage failure reasoning tree as constraint context into the large language model; The structured entities (nodes at each level and their attribute information), hierarchical relationships (parent-child subordinate relationships and sibling node parallel relationships), node-embedded rules (damage and failure knowledge rules bound to each node), and the tree-topology-fixed thought chain reasoning path (the complete reasoning path sequence from the root node to the leaf node) contained in the multi-level damage and failure reasoning tree are all input into the large language model as constraint context. The purpose is to define a clear reasoning track for the large language model, so that it can only operate within the knowledge scope defined by the tree topology structure in subsequent reasoning processes, and cannot call on external irrelevant knowledge or introduce unverified engineering experience, thereby controlling the knowledge boundary and logical order of reasoning activities from the source stage.

[0076] Step S302: The large language model traverses the damage failure classification nodes of the multi-level damage failure reasoning tree layer by layer from top to bottom along the thought chain reasoning path of the tree topology hierarchy, and deduces the damage failure step by step in a chain manner according to the thought chain reasoning path of pressure equipment, damage failure category, subdivided damage mode, and calculation of damage failure probability. The large language model follows a tree-like topological hierarchy of reasoning paths, traversing each level of damage and failure classification nodes from top to bottom in a fixed hierarchical order: private root node (pressure-bearing equipment), first intermediate branch node (damage and failure category), second intermediate branch node (detailed damage pattern), and leaf node (failure probability calculation). Reasoning can only proceed to the next level after all candidate nodes in the current level have completed their reasoning and output results. The damage and failure categories and their corresponding detailed damage patterns are output level by level for each branch node, forming a complete chain-like derivation record. The purpose is to force the large language model to perform step-by-step derivations according to a preset engineering logic order by changing the reasoning path from dynamically generated to hierarchically fixed. This avoids logical jumps or path deviations that might occur if the model freely chooses its reasoning order, ensuring the transparency and traceability of the reasoning process.

[0077] Step S303: During the hierarchical derivation process, the damage failure knowledge rules and the damage failure constraint parameters are synchronously called to verify the reasoning process step by step, and the reasoning branches generated during the reasoning process are verified according to industry standard clauses and working condition numerical boundary conditions to complete the topology constraint enhancement.

[0078] During the hierarchical derivation process, damage and failure knowledge rules and damage and failure constraint parameters are synchronously invoked to verify the intermediate conclusions of each step of the reasoning process. Furthermore, the generated reasoning branches are validated for compliance based on industry standard clauses and working condition boundary conditions, intercepting any abnormal reasoning branches that do not conform to engineering specifications or exceed working condition boundary conditions. Step-by-step verification is embedded within the reasoning process and is executed alternately with model reasoning. After each node's judgment is completed, rules and parameters are invoked for verification. Branch verification is performed after the reasoning branch is completed, and it is released only after confirming that there are no conflicts or contradictions. These two verifications form a dual verification mechanism of "process and end result." This dual verification ensures that the reasoning conclusions at each level are based on evidence and conform to engineering safety boundaries, thus enhancing topological constraints. The purpose is to establish a real-time quality monitoring mechanism during the reasoning process, preventing the large language model from generating conclusions that do not meet standard requirements or exceed working condition boundaries, ensuring the safety and compliance of the reasoning results in engineering, and providing validated and reliable input for subsequent failure probability calculations.

[0079] For further embodiments, please refer to Figure 5After retrieving all required parameters using the damage mode output by the leaf node as an index, the damage failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage failure probability. This process includes: Step S3001: Using the subdivided damage modes output by each leaf node as indexes, in the shared node, the degradation rate coefficient and non-destructive testing reduction coefficient corresponding to each damage mode are retrieved in parallel, and the damage co-coupling amplification index is obtained by combining and matching all identified damage types; the cumulative service time, operating condition fluctuation deviation coefficient and detection data missingness coefficient are retrieved from the private link; the baseline inherent failure probability and weight coefficient are retrieved from the shared node. Step S3002: Calculate the multi-damage independent degradation cumulative term based on the actual number of detected damage types, the degradation rate coefficient, and the cumulative service time; For each specific damage mode actually existing on the equipment, its degradation rate coefficient (such as the annual corrosion rate of uniform corrosion, the propagation rate of fatigue cracks, etc.) is multiplied by the cumulative service time to obtain the cumulative degradation amount of that damage over the service life. Then, the cumulative degradation amounts of all damages are summed to obtain the overall independent degradation cumulative term of the equipment. This value reflects the overall damage degree under the independent action of various types of damage, and retains the individual differences of various types of damage. The damage mode with a larger degradation rate coefficient contributes more to the total degradation amount.

[0080] Step S3003: Perform an exponentiation operation on the damage synergistic coupling amplification index and the multi-damage independent degradation accumulation term to obtain the multi-damage coupling amplification factor; perform a multiplication operation on the non-destructive testing reduction coefficient to obtain the hierarchical targeted inspection correction factor. The sum of the multi-damage independent degradation accumulation term calculated in step S3002 and the numerical value one is then powered with the damage synergistic coupling amplification index to obtain the multi-damage coupling amplification factor. When multiple types of damage exist simultaneously in a device, synergistic effects may occur between these types of damage. For example, stress concentration at the bottom of a corrosion pit can accelerate fatigue crack initiation, and high-temperature environments can accelerate corrosion rates, leading to a nonlinear amplification of failure risk. The damage synergistic coupling amplification index is used to quantify this nonlinear coupling amplification effect. When only a single type of damage exists, the index degenerates to a numerical value of one, with no coupling amplification. When multiple types of damage exist, the index is greater than a numerical value of one, exponentially amplifying the degradation accumulation term. The more types of damage and the more complex the synergistic effect, the more significant the amplification effect.

[0081] Furthermore, by multiplying the nondestructive testing reduction coefficients, a stratified targeted testing correction factor is obtained.

[0082] The non-destructive testing (NDT) reduction factors corresponding to each sub-damage mode are multiplied together to obtain a stratified targeted testing correction factor. Different types of damage have different sensitivities to testing methods. For stress corrosion cracking, ultrasonic phased array testing can achieve a high detection rate and a significant reduction in failure probability, with a smaller reduction factor. Conventional ultrasonic testing may have a lower detection rate and a limited reduction in failure probability, with a larger reduction factor. Each sub-damage mode has its industry-standard recommended testing method and corresponding reduction factor, ranging from zero to one. The closer the value of λ is to zero, the higher the detection effectiveness and the more significant the reduction effect; the closer the value is to one, the lower the detection effectiveness and the more limited the reduction effect. Multiplying these reduction factors together achieves a comprehensive quantification of the testing effectiveness when multiple types of damage coexist. The more types of damage there are, the more the independent detection effectiveness of each item is superimposed, and the multiplicative result shows a cumulative reduction effect. When a certain type of damage is not effectively detected, its reduction factor approaches one and produces no reduction effect.

[0083] Step S3004: After weighted summation of the working condition fluctuation deviation coefficient and its corresponding first weight coefficient, the detection data missingness coefficient and its corresponding second weight coefficient, add the summation to the value to obtain the uncertainty correction factor; The uncertainty correction coefficient is obtained by weighting and summing the operating condition fluctuation deviation coefficient and its corresponding first weighting coefficient, and the detection data missingness coefficient and its corresponding second weighting coefficient. The sum of this correction coefficient and a single value is used as the uncertainty correction factor. There are two sources of uncertainty in the assessment process: the operating condition fluctuation deviation coefficient reflects the degree to which the actual operating temperature or pressure deviates from the design operating conditions; the greater the deviation, the greater the possibility that the actual risk is higher than the theoretically calculated value. The detection data missingness coefficient reflects the completeness of historical inspection records; the more data missing, the greater the uncertainty of the assessment. The two types of deviations are weighted and summed using preset weighting coefficients, and then the single value is added to obtain the overall uncertainty correction factor. When both types of deviations are zero, the factor is a single value, indicating no additional uncertainty correction; the greater the deviation, the larger the factor is than a single value, and the failure probability is adjusted upwards accordingly.

[0084] Step S3005: Perform a product operation based on the baseline inherent failure probability, the multi-damage coupling amplification factor, the hierarchical targeted inspection correction factor, and the uncertainty correction factor to obtain the damage failure probability of the pressure-bearing equipment.

[0085] The baseline inherent failure probability is obtained from tables in standard specifications based on equipment type, material, and design parameters. It represents the baseline risk level of this type of equipment under ideal conditions and serves as the starting point for failure probability calculation. The multi-damage coupling amplification factor amplifies the baseline probability, reflecting the nonlinear coupling effect when multiple damages coexist. The hierarchical targeted inspection correction factor reduces the amplified probability, reflecting the effect of effective inspection on reducing failure risk. The uncertainty correction factor further adjusts the uncertainty of the final result, reflecting the additional risks caused by operating condition fluctuations and data gaps. The baseline inherent failure probability of pressure equipment is calculated by multiplying the baseline inherent failure probability, the multi-damage coupling amplification factor, the hierarchical targeted inspection correction factor, and the uncertainty correction factor sequentially, forming a complete quantitative model of damage failure probability. This realizes a complete quantitative chain starting from the baseline probability, through multi-damage coupling amplification, inspection effectiveness reduction, and uncertainty correction.

[0086] In some specific embodiments, the expression for the damage failure probability calculation model is: in, The probability of damage failure of pressure equipment is a value that characterizes the final failure probability of pressure equipment under current operating conditions. The baseline inherent failure probability of the equipment is obtained by looking up tables in industry standards and specifications (such as API 581 and GB / T 26610) based on the equipment type, material, and design parameters. Similar equipment shares the same baseline value, which belongs to the knowledge rules of damage and failure. The number of subdivided damage types actually detected is based on the total number of damage mode types that are still present on the equipment and confirmed by the most recent valid in-service inspection. It is a damage failure constraint parameter for a single pressure equipment. For the first The degradation rate coefficient of subdivided damage is obtained by actual calculation based on the actual operating conditions of the equipment (medium, temperature, pressure) and detection data (such as the annual corrosion rate of uniform corrosion, the propagation rate of fatigue cracks, etc.), and belongs to the damage failure constraint parameters of a single pressure-bearing equipment. Cumulative service duration represents the actual number of years of operation from the date the equipment was put into use to the time of assessment. It is a damage and failure constraint parameter for a single pressure-bearing device. This is a cumulative term of independent degradation from multiple injuries, characterizing the overall degree of damage degradation under the independent action of various types of damage; The damage co-coupling amplification index is given by industry standards based on the failure category and the type of coexisting damage combination. It is used to quantify the nonlinear coupling amplification effect when multiple damages coexist, and belongs to the damage failure knowledge rules. For the first The non-destructive testing reduction factor corresponding to the sub-type damage is uniformly specified by industry standards for each sub-type damage mode and its recommended testing method. The value range is between 0 and 1. It is used to quantify the effect of effective testing on reducing the probability of failure and belongs to the damage failure knowledge rules. To provide a stratified targeted test correction factor, characterizing the cumulative reduction effect of test effectiveness when multiple damages coexist; It is the first weighting coefficient corresponding to the operating condition fluctuation deviation coefficient, which is fixed by industry standards and is used to weigh the weight of the operating condition fluctuation deviation in the overall uncertainty. It belongs to the damage and failure knowledge rules. The second weighting coefficient corresponding to the missing data coefficient is fixed by industry standards and is used to weigh the weight of missing data in the overall uncertainty. It belongs to the damage and failure knowledge rules. Operating condition fluctuation deviation coefficient (subscript) (Represents the actual operating temperature of the equipment). It represents the degree of deviation of key process parameters such as temperature, pressure, and flow rate, based on the actual operating temperature of the equipment. It belongs to the damage and failure constraint parameters of a single pressure equipment. The missing data coefficient is calculated based on the completeness of the equipment's historical inspection records (wall thickness, cracks, corrosion data), and is a damage failure constraint parameter for a single pressure-bearing device.

[0087] Step S40: Based on the failure probability level corresponding to the damage failure probability and the failure consequence level corresponding to the damage failure constraint parameter, determine the risk level of the target pressure-bearing equipment and generate a risk assessment conclusion. Based on the tree-structured topology and the fixed thought chain reasoning path, a forced step-by-step deduction constraint is formed, which enables the large language model to perform hierarchical reasoning according to the tree-structured topology. It outputs the damage failure category and its corresponding sub-damage mode of each branch node level by level. After retrieving all the required parameters by using the damage mode output by the leaf node as an index, the damage failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage failure probability. Then, based on the failure probability level corresponding to the damage failure probability and the failure consequence level corresponding to the damage failure constraint parameters, the risk level of the target pressure equipment is determined, and a risk assessment conclusion is generated.

[0088] In some embodiments, please refer to Figure 6 The steps of determining the risk level of the target pressure-bearing equipment and generating a risk assessment conclusion based on the failure probability level corresponding to the damage failure probability and the failure consequence level corresponding to the damage failure constraint parameter include: Step S401: Obtain the working condition boundary conditions, medium component limitation range, and medium concentration limitation range from the damage failure constraint parameters; The operating condition numerical boundary conditions, medium component limitation ranges, and medium concentration limitation ranges are extracted from the damage and failure constraint parameters. The operating condition numerical boundary conditions are the limit values ​​and rated ranges of key process parameters such as temperature, pressure, and flow rate during actual equipment operation. The medium component limitation ranges refer to the types and content ranges of chemical components contained in the internal medium of the equipment (such as hydrogen, hydrogen sulfide, carbon dioxide, chloride ions, etc.). The medium concentration limitation ranges refer to the specific concentration ranges and fluctuation ranges of various corrosive components. This step provides basic input data for subsequent determination of the failure consequence level.

[0089] Step S402: Determine the failure consequence level of the target pressure-bearing equipment based on the operating condition boundary conditions, the defined range of medium components, and the defined range of medium concentration. Based on the operating condition boundary conditions, medium component limits, and medium concentration limits obtained in step S401, the failure consequence level of the target pressure-bearing equipment is determined. The failure consequence level refers to the severity of the potential consequences, such as personal injury, property damage, environmental impact, and production interruption, that may result from equipment failure. Specifically, the operating condition boundary conditions (temperature, pressure) affect the leakage rate and diffusion behavior; the medium component determines the toxicity (e.g., hydrogen sulfide is highly toxic) and flammability (e.g., hydrogen is flammable); and the medium concentration affects the hazard intensity of the leaked medium. In RBI (Risk-Based Inspection) engineering practice, the numerical boundary conditions of the operating conditions, the limited range of medium components, and the limited range of medium concentration are the core input factors for failure consequence assessment. By using the failure consequence calculation methods specified in standards such as API RP 581 or GB / T 26610.5, the calculated quantitative values ​​of failure consequences are mapped to a preset consequence level range (such as consequence level 1 to consequence level 5, where the higher the consequence level, the more severe the possible consequences such as personal injury, property damage, environmental impact, and production interruption). This results in a discretized failure consequence level, which is used for subsequent cross-location of the risk matrix.

[0090] Step S403: Map the damage failure probability to a preset failure probability level range to determine the failure probability level of the target pressure-bearing equipment; The damage failure probability value calculated by the failure probability calculation model built into the leaf node is mapped according to a preset probability level range to determine the failure probability level of the target pressure-bearing equipment. The failure probability value is typically a continuous value between 0 and 1. It needs to be mapped to a discrete level (e.g., failure level 1 to failure level 5, where higher failure levels correspond to higher damage failure probabilities) according to industry standards or engineering practices. This converts the quantitatively calculated continuous probability value into a discrete level that is easy to locate in the risk matrix in engineering, making the risk level determination process standardized and operable. The failure probability level range corresponding to failure level 1 can be... The failure probability level range corresponding to the level 2 failure can be: The failure probability level range corresponding to the third failure level can be: The failure probability level range corresponding to the fourth failure level can be: The failure probability level range corresponding to the fifth failure level can be: Those skilled in the art can adjust the failure probability level range as needed according to the actual application scenario, without making any limitations here.

[0091] Step S404: Based on the failure probability level and the failure consequence level, locate the risk coordinates of the target pressure-bearing equipment in a preset risk matrix, wherein the risk matrix is ​​a two-dimensional risk assessment table constructed with the failure probability level as the vertical axis and the failure consequence level as the horizontal axis, and each intersection point in the two-dimensional risk assessment table corresponds to a risk level. Using the failure probability level as the vertical axis (rows) and the failure consequence level as the horizontal axis (columns), cross-locate the data within a pre-defined risk matrix to obtain a unique risk coordinate corresponding to the target pressure-bearing equipment. The risk matrix is ​​a pre-defined two-dimensional risk assessment table, where each intersection point corresponds to a risk level (e.g., low risk, medium risk, high risk, extremely high risk). The vertical axis reflects the probability of equipment failure, and the horizontal axis reflects the severity of the consequences of equipment failure; combining these two aspects determines the overall risk level of the equipment. Step S404 cross-locates the failure consequence level determined in step S402 and the failure probability level determined in step S403 within the risk matrix, forming a unique coordinate point, thus transforming the two independent risk levels into a one-dimensional comprehensive risk.

[0092] Step S405: Take the risk level corresponding to the risk coordinate as the risk level of the target pressure-bearing equipment, and generate a risk assessment conclusion that includes the risk level determination result and the failure consequence level determination result.

[0093] The risk level corresponding to the risk coordinate in step S404 is taken as the risk level of the target pressure equipment, and a risk assessment conclusion including the risk level determination result and the failure consequence level determination result is generated. This conclusion provides inspection personnel with a clear basis for decision-making: different risk levels correspond to different inspection strategies and maintenance priorities (e.g., extremely high risk requires immediate inspection measures, high risk should be prioritized for inspection, medium risk should be inspected according to the normal cycle, and low risk can appropriately extend the inspection cycle). At the same time, the failure consequence level determination result also provides additional reference for the subsequent inspection strategy formulation (e.g., equipment with a high failure consequence level should be given sufficient attention even if the probability of damage failure is low).

[0094] In some embodiments, the damage failure probability and its corresponding failure likelihood level of the pressure equipment can be obtained based on equipment parameters (basic information), operating conditions (equipment design / operating temperature and pressure, vibration / temperature fluctuations in the production environment), medium characteristics (medium, composition, phase, corrosive medium content, etc.), service records (commissioning time, maintenance records), and the damage failure probability calculation model built into each leaf node. Based on the equipment volume, medium characteristics (medium, composition, phase, flammability, toxicity, corrosivity, etc.), and the operating condition boundary conditions, medium component limitation range, and medium concentration limitation range in the damage failure constraint parameters, the impact area and economic consequences of the pressure equipment failure can be obtained, and the failure consequence level of the pressure equipment can be obtained. The risk level of the target pressure equipment can be comprehensively determined by a two-dimensional risk matrix based on the failure likelihood level and the failure consequence level, and a risk assessment conclusion can be generated. Step S50: Archive the reasoning path of the thought chain, the industry standard clauses corresponding to each branch node, the calculation basis of the probability of damage failure, and the judgment basis of the risk assessment conclusion to form a traceable file and output a natural description of the risk assessment of the pressure equipment.

[0095] Based on the failure probability level corresponding to the damage failure probability and the failure consequence level corresponding to the damage failure constraint parameters, the risk level of the target pressure equipment is determined. After generating the risk assessment conclusion, the reasoning path of the thought chain, the industry standard clauses corresponding to each branch node, the calculation basis of the damage failure probability, and the judgment basis of the risk assessment conclusion are archived to form a traceable file and output a natural description of the risk assessment of the pressure equipment.

[0096] In some embodiments, please refer to Figure 7 The steps involved in archiving the thought process reasoning path, the industry standard clauses corresponding to each branch node, the calculation basis for the probability of damage and failure, and the judgment basis for the risk assessment conclusion, to form a traceable archive and output a natural description of the risk assessment of pressure equipment, include: Step S501: Collect the reasoning records generated at each level during the hierarchical reasoning process of the large language model along the tree topology hierarchy to form the thought chain reasoning path; To transform the internal reasoning process of the large language model from a "black box" into a traceable and transparent chain, laying the data foundation for the subsequent generation of traceable archives, we collect reasoning records generated at each level during the hierarchical reasoning process of the large language model along its tree-like topology, forming a thought chain reasoning path. During the hierarchical reasoning process, each level (such as the damage / failure category determination layer, the subdivided damage pattern recognition layer, etc.) generates intermediate reasoning records, including the constraints input at the current level, the invoked knowledge rules, and the intermediate conclusions derived from the reasoning. These records are collected and linked together in the order in which the reasoning occurs, thus forming a complete thought chain reasoning path, fully recording the entire reasoning trajectory of the model from the starting node to the leaf node.

[0097] Step S502: Collect the original text of the industry standard clauses and their clause numbers called by each branch node during the reasoning process, and form a standard clause call record; This provides clear engineering specifications for each subsequent conclusion, ensuring that the output of the large language model is based on evidence and verifiable, avoiding unfounded "illusionary" outputs. The system collects the original text and clause numbers of industry standard clauses invoked by each branch node during the inference process, forming a standard clause invocation record. During the inference process, whenever a major category or sub-category of damage failure is identified, the large language model invokes the corresponding industry standard clauses (such as the judgment rules for a certain type of damage mode) as the basis for judgment. The system automatically collects the original text and numbers of these invoked clauses, forming a complete invocation record.

[0098] Step S503: Collect the calculation basis for the damage failure probability and the judgment basis for the risk assessment conclusion, and form a conclusion support record; To comprehensively record the basis for conclusions from both the calculation and judgment processes, support the verifiability of risk assessment conclusions, and enable inspectors to verify the source of each value and the logic of each judgment during review, the system collects the calculation basis for damage failure probability and the judgment basis for risk assessment conclusions, forming a conclusion support record. The calculation basis for damage failure probability includes the source and value path of the degradation rate coefficient for each damage, the specific value of the cumulative service time, the value basis of the damage synergistic amplification index, the standard source and value of each non-destructive testing reduction coefficient, and a record of the calculation process for each parameter. The judgment basis for risk assessment conclusions includes the cross-location process of failure probability level and failure consequence level in the risk matrix, the standard basis for risk level judgment, and the logical link in the formation of the final conclusion. The system automatically collects all the above information to form a conclusion support record.

[0099] Step S504: The reasoning path of the thought chain, the standard clause call record and the conclusion support record are structurally associated according to the reasoning sequence to generate a traceable archive of the entire reasoning process; The scattered reasoning paths, standard clause invocation records, and conclusion support records are integrated into a structured archive with clear logical connections according to the reasoning sequence. This allows examiners to fully reproduce each step of the reasoning process from initial input to final conclusion, achieving full traceability. The reasoning paths, standard clause invocation records, and conclusion support records are structured and linked according to the reasoning sequence, generating a traceable archive of the entire reasoning process. These three types of information are organized according to the chronological order of reasoning occurrence: the first reasoning step, the standard clauses invoked, and the calculation or judgment conclusions they support, establishing logical connections between them to form a complete archive arranged chronologically with clear causal relationships between each information segment.

[0100] Step S505: Based on the traceable archive, output a risk assessment natural description in natural language form, including a summary of the reasoning chain, key judgment criteria, and risk assessment conclusions.

[0101] To transform the structured data in traceable archives into easily readable text for inspectors, this approach preserves traceable technical details while enhancing report readability through natural language. This allows inspectors to understand the formation process and basis of assessment conclusions without needing to manually parse the structured data. Based on the traceable archives, a natural language description of the risk assessment is output, including a summary of the reasoning chain path, key judgment criteria, and risk assessment conclusions. The reasoning chain summary is a concise overview of the thought chain reasoning path in step S501, presenting the core reasoning steps in a clear manner rather than listing all intermediate variables verbatim. The key judgment criteria are core summaries of the standard clause citation records and conclusion support records in steps S502 and S503, highlighting the key standard clauses and main calculation bases supporting the final conclusion. The final output is a complete and readable risk assessment natural language report.

[0102] As can be seen from the above embodiments, compared with the prior art, this application addresses the problems of insufficient evaluation consistency caused by the high dependence of RBI methods on human experience in engineering implementation, the difficulty in systematically reusing domain knowledge due to its fragmented nature, the limited knowledge integration and engineering interpretation capabilities of commercial software, and the existence of knowledge gaps, illusions, and poor interpretability in general large language models under strongly standardized scenarios. This application includes, but is not limited to, the following beneficial effects: Firstly, this application addresses the issue of excessive reliance on human experience in the risk assessment process, resulting in insufficient consistency and traceability. It establishes a rigid, tree-structured topology for the reasoning path, creating a mandatory step-by-step deduction constraint. This ensures the large language model strictly adheres to the hierarchical order of pressure-bearing equipment, damage / failure categories, subdivided damage modes, and leaf nodes, transforming the judgment process dependent on personal experience into a standardized, deterministic reasoning process. The reasoning conclusions of each branch node are based on industry standard clauses and knowledge rules uniformly bound within the shared node, avoiding assessment biases caused by differences in knowledge background and experience among different inspectors. Furthermore, by comprehensively archiving the reasoning path, standard clause citation records, and calculation / judgment basis through traceable documentation, the formation process of the assessment conclusions becomes verifiable and traceable, significantly improving the consistency and transparency of RBI analysis.

[0103] Secondly, this application addresses the issue of fragmented and unsystematic reuse of domain knowledge by constructing a knowledge graph for pressure equipment. This graph unifies the damage and failure knowledge rules from standards and specifications such as GB / T 26610 (Guidelines for Risk-Based Inspection of Pressure Equipment Systems), GB / T 30579 (Damage Pattern Recognition of Pressure Equipment), and API RP 571 (Damage Mechanisms of Fixed Equipment in the Refining Industry) into a globally shared node. Simultaneously, it structures and organizes historical RBI project reports, inspection records, and typical engineering cases to establish a case knowledge base, achieving systematic integration of fragmented knowledge. All pressure equipment shares the same set of knowledge rules within the shared node. New equipment only needs to configure its private constraint parameters to reuse the complete knowledge system, effectively reducing the marginal cost of knowledge maintenance and improving the reuse efficiency of engineering experience and historical cases.

[0104] Thirdly, addressing the limited knowledge integration and engineering interpretation capabilities of existing RBI software, this application uses a large language model as the inference engine. During the inference process, it synchronously calls damage and failure knowledge rules and constraint parameters for step-by-step verification, and performs compliance verification on the inference branches based on industry standard clauses and working condition boundary conditions, thus realizing a systematic expression of the engineering logic behind the standard clauses. By completely archiving the original text and number of the standard clauses called by each branch node through traceable archives, each inference conclusion has a clear engineering basis, significantly improving the engineering interpretation capability. At the same time, by establishing independent association links between private root nodes and each shared node, it realizes the separate management of general knowledge rules and individual constraint parameters, taking into account both the uniformity of the knowledge system and the flexibility of working condition adaptation, overcoming the limitations of existing commercial software that emphasizes process calculation and neglects knowledge integration and engineering interpretation.

[0105] Fourth, addressing the insufficient applicability of general-purpose large language models in strictly regulated engineering scenarios, this application uses a multi-level damage and failure reasoning tree as a structured retrieval enhancement index, thus limiting the knowledge sources and analytical scope of the large language model. It replaces the reasoning links dynamically generated by prompt words with a tree-topology-fixed hierarchical thought chain reasoning path, thereby suppressing model reasoning deviation and "illusion" risks from a mechanism perspective. Through a "dual-engine" architecture where the rule engine and the knowledge-enhanced large language model work together, knowledge rules and constraint parameters are synchronously invoked for step-by-step verification during the reasoning process. Furthermore, compliance checks are performed on the reasoning branches based on industry standard clauses and working condition boundary conditions. This forms a triple-credible reasoning mechanism: "domain knowledge graph constrains knowledge sources, fixed thought chain constrains reasoning paths, and rule engine constrains output results." This transforms the large language model from a potentially "illusionary" free text generator into a credible reasoning tool that strictly follows the hierarchical paths fixed by the engineering knowledge graph to perform forced step-by-step deductions, enabling the reliable application of the large language model in the high-safety, strictly regulated engineering scenario of pressure equipment RBI.

[0106] Furthermore, while maintaining the advantages of semantic understanding and natural language interaction of large language models, this application strictly constrains its reasoning behavior within the boundaries defined by engineering specifications and knowledge rules. This effectively suppresses the reasoning deviation and "illusion" risk of general large language models in strongly regulated engineering scenarios, and realizes the standardization of the evaluation process, the transparency of the reasoning path, and the traceability of the decision-making basis. It provides an intelligent risk assessment scheme for pressure equipment RBI that combines knowledge integrity, engineering constraints, and interpretability of results.

[0107] Please see Figure 8This application provides a pressure equipment risk assessment device based on a knowledge graph-enhanced large language model, which includes a reasoning tree generation module 1100, a node configuration module 1200, a hierarchical reasoning module 1300, a risk level determination module 1400, and a traceable archive archiving module 1500, all for the purposes of this application. The inference tree generation module 1100 is configured to call the knowledge graph of pressure equipment, using each pressure equipment and its corresponding risk assessment association data as matching conditions, and perform layer-by-layer matching and filtering in the multi-level damage and failure classification structure of the knowledge graph to extract a subset of damage and failure classification nodes that are suitable for the pressure equipment and have a tree-like topology, and generate a multi-level damage and failure inference tree corresponding to the pressure equipment; the node configuration module 1200 is configured to determine the root node of the subset of damage and failure classification nodes as a private root node, and determine its first intermediate branch node, second intermediate branch node, and leaf node as shared nodes that can be reused by all pressure equipment, and uniformly bind damage and failure knowledge rules within all shared nodes; establish mutually independent association links between the private root node and each shared node, and store damage and failure constraint parameters that are only suitable for the corresponding pressure equipment operating conditions in an isolated manner in each association link; the hierarchical inference module 1300 is configured to be based on a tree-like topology hierarchy. A fixed thought chain reasoning path forms a mandatory step-by-step deduction constraint, enabling the large language model to perform hierarchical reasoning according to a tree-like topology. It outputs the major damage and failure categories and their corresponding sub-damage modes for each branch node level by level. After retrieving all required parameters using the damage modes output by the leaf nodes as indexes, the damage and failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage and failure probability. The risk level determination module 1400 is configured to determine the risk level of the target pressure-bearing equipment based on the failure probability corresponding to the failure likelihood level and the failure consequence level corresponding to the damage and failure constraint parameters, generating a risk assessment conclusion. The traceable archive module 1500 is configured to archive the thought chain reasoning path, the industry standard clauses corresponding to each branch node, the calculation basis of the damage and failure probability, and the determination basis of the risk assessment conclusion, forming a traceable archive and outputting a natural description of the risk assessment of the pressure-bearing equipment.

[0108] Based on any embodiment of this application, please refer to Figure 9 Another embodiment of this application also provides an electronic device, which can be implemented by a computer device, such as... Figure 9The diagram shows the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database stores control information sequences. When the computer-readable instructions are executed by the processor, they enable the processor to implement a pressure equipment risk assessment method based on a knowledge graph-enhanced large language model. The processor provides computational and control capabilities to support the operation of the entire computer device. The memory stores computer-readable instructions, which, when executed by the processor, enable the processor to execute the pressure equipment risk assessment method based on a knowledge graph-enhanced large language model of this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0109] In this embodiment, the processor is used to execute... Figure 8 The specific functions of each module are defined within the device, and the memory stores the program code and various data required to execute these modules. A network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules in the pressure equipment risk assessment device based on a knowledge graph-enhanced large language model of this application. The server can call the server's program code and data to execute the functions of all modules.

[0110] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the pressure equipment risk assessment method based on a knowledge graph-enhanced large language model as described in any embodiment of this application.

[0111] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the pressure equipment risk assessment method based on a knowledge graph-enhanced large language model as described in any embodiment of this application.

[0112] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0113] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A risk assessment method for pressure equipment based on a knowledge graph-enhanced large language model, characterized in that, include: The knowledge graph of pressure equipment is invoked, and each pressure equipment and its corresponding risk assessment data are used as matching conditions. The multi-level damage and failure classification structure of the knowledge graph is matched and filtered layer by layer to extract the subset of damage and failure classification nodes that are suitable for the pressure equipment and have a tree topology, and generate a multi-level damage and failure reasoning tree corresponding to the pressure equipment. The root node in the subset of damage and failure classification nodes is determined as a private root node, and its first intermediate branch node, second intermediate branch node, and leaf node are determined as shared nodes that can be reused by all pressure-bearing equipment. Damage and failure knowledge rules are uniformly bound within all shared nodes. Independent association links are established between the private root node and each shared node, and each association link stores damage and failure constraint parameters that are only suitable for the corresponding pressure-bearing equipment operating conditions in an isolated manner. Based on the tree-topology hierarchical fixed thinking chain reasoning path, a forced step-by-step deduction constraint is formed, which enables the large language model to perform hierarchical reasoning according to the tree-topology hierarchical order, outputting the damage failure category and its corresponding subdivided damage mode of each branch node level by level. After retrieving all the required parameters by using the damage mode output by the leaf node as an index, the damage failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage failure probability. Based on the failure probability level corresponding to the damage failure probability and the failure consequence level corresponding to the damage failure constraint parameter, the risk level of the target pressure equipment is determined, and a risk assessment conclusion is generated. The archived thought chain reasoning path, industry standard clauses corresponding to each branch node, calculation basis for damage failure probability, and judgment basis for risk assessment conclusion form a traceable archive, outputting a natural description of the risk assessment of pressure equipment.

2. The method for risk assessment of pressure equipment based on a knowledge graph-enhanced large language model according to claim 1, characterized in that, The damage failure classification node subset includes a private root node representing the pressure-bearing equipment, a first intermediate branch node representing the major damage failure category, a second intermediate branch node representing the subdivided damage mode, and a leaf node representing the damage failure probability. The pressure-bearing equipment includes a complete set of petrochemical plant pressure vessels and pressure pipelines, wherein the complete set of petrochemical plant pressure vessels includes a reaction vessel, a heat exchange vessel, a separation vessel and a storage vessel. The risk assessment data includes cumulative service time, number of actual detected damage types, degradation rate coefficient of each damage, operating condition fluctuation deviation coefficient, detection data missingness coefficient, material mechanical property parameters, and corrosion resistance parameter. The damage failure knowledge rules include corrosion mechanisms, industry standard clauses, damage failure probability calculation models, and detection and maintenance rules corresponding to each subdivided damage mode. The damage and failure constraint parameters include the operating condition boundary conditions, the range of medium composition limits, and the range of medium concentration limits.

3. The method for risk assessment of pressure equipment based on a knowledge graph-enhanced large language model according to claim 1, characterized in that, The steps of matching and filtering layer by layer in the multi-level damage and failure classification structure of the knowledge graph to extract a subset of damage and failure classification nodes that are suitable for the pressure-bearing equipment and have a tree-like topology, and generating a multi-level damage and failure inference tree corresponding to the pressure-bearing equipment, include: Feature extraction is performed on the pressure-bearing equipment and its corresponding risk assessment data to obtain a feature vector related to the risk assessment target of the pressure-bearing equipment. Based on the feature vector, locate the initial damage failure classification node that matches the risk assessment target in the knowledge graph of the pressure equipment; Starting from the initial damage failure classification node, following the multi-level damage failure classification structure in the pressure equipment knowledge graph, depth-first retrieval is performed level by level according to the preset thought chain reasoning path to extract the subset of damage failure classification nodes that constitute the multi-level damage failure reasoning tree; The subset of damage and failure classification nodes is organized into a tree topology according to hierarchical relationships to generate a multi-level damage and failure reasoning tree adapted to the pressure-bearing equipment.

4. The method for risk assessment of pressure equipment based on a knowledge graph-enhanced large language model according to claim 1, characterized in that, Based on a tree-structured topological hierarchy-based reasoning path, a mandatory step-by-step deduction constraint is formed, enabling the large language model to perform hierarchical reasoning according to the tree-structured topological hierarchy, and outputting the damage and failure categories and their corresponding sub-damage modes of each branch node level by level, including: The structured entities, hierarchical relationships, node-embedded rules, and thought chain reasoning paths of the multi-level damage failure reasoning tree are input as constraint contexts into the large language model. The large language model traverses the damage and failure classification nodes of the multi-level damage and failure reasoning tree layer by layer from top to bottom along the thought chain reasoning path of the pressure-bearing equipment, damage and failure category, subdivided damage mode, and calculation of damage and failure probability in a step-by-step chain derivation. During the hierarchical derivation process, the damage failure knowledge rules and the damage failure constraint parameters are synchronously invoked to verify the reasoning process step by step. The reasoning branches generated during the reasoning process are verified according to industry standard clauses and working condition numerical boundary conditions to complete the topology constraint enhancement.

5. The method for risk assessment of pressure equipment based on a knowledge graph-enhanced large language model according to claim 2, characterized in that, After retrieving all required parameters using the damage mode output by the leaf node as an index, the damage failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage failure probability. This includes the following steps: Using the subdivided damage modes output by each leaf node as an index, the degradation rate coefficient and non-destructive testing reduction coefficient corresponding to each damage mode are retrieved in parallel in the shared node, and the damage co-coupling amplification index is obtained by combining and matching all identified damage types; the cumulative service time, operating condition fluctuation deviation coefficient and detection data missingness coefficient are retrieved from the private link; the baseline inherent failure probability and weight coefficient are retrieved from the shared node. Based on the number of actual detected damage types, the degradation rate coefficient, and the cumulative service time, calculate the multi-damage independent degradation cumulative term; The multi-damage coupling amplification factor is obtained by exponentiation of the damage synergistic coupling amplification index and the multi-damage independent degradation accumulation term; the hierarchical targeted inspection correction factor is obtained by multiplying the non-destructive testing reduction coefficient. The uncertainty correction factor is obtained by weighting and summing the working condition fluctuation deviation coefficient and its corresponding first weight coefficient, the detection data missingness coefficient and its corresponding second weight coefficient, and then adding the sum to the value. The damage failure probability of the pressure equipment is obtained by multiplying the baseline inherent failure probability, the multi-damage coupling amplification factor, the hierarchical targeted inspection correction factor, and the uncertainty correction factor.

6. The method for risk assessment of pressure equipment based on a knowledge graph-enhanced large language model according to claim 1, characterized in that, The steps for determining the risk level of the target pressure-bearing equipment and generating a risk assessment conclusion based on the failure probability level corresponding to the damage failure probability and the failure consequence level corresponding to the damage failure constraint parameters include: Obtain the working condition boundary conditions, medium component limitation range, and medium concentration limitation range from the damage failure constraint parameters; The failure consequence level of the target pressure-bearing equipment is determined based on the numerical boundary conditions of the operating conditions, the defined range of the medium components, and the defined range of the medium concentration. The failure probability is mapped to a preset failure probability level range to determine the failure probability level of the target pressure-bearing equipment; Based on the failure probability level and the failure consequence level, the risk coordinates of the target pressure-bearing equipment are located in a preset risk matrix. The risk matrix is ​​a two-dimensional risk assessment table constructed with the failure probability level as the vertical axis and the failure consequence level as the horizontal axis. Each intersection point in the two-dimensional risk assessment table corresponds to a risk level. The risk level corresponding to the risk coordinate is taken as the risk level of the target pressure-bearing equipment, and a risk assessment conclusion including the risk level determination result and the failure consequence level determination result is generated.

7. The method for risk assessment of pressure equipment based on a knowledge graph-enhanced large language model according to any one of claims 1 to 6, characterized in that, The steps involved in archiving the thought process reasoning path, the industry standard clauses corresponding to each branch node, the calculation basis for the probability of damage and failure, and the judgment basis for the risk assessment conclusion, to form a traceable archive and output a natural description of the risk assessment of pressure equipment, include: Collect the reasoning records generated at each level during the hierarchical reasoning process of the large language model along the tree topology, and form the thought chain reasoning path; Collect the original text and clause number of the industry standard clauses called by each branch node during the reasoning process, and form a standard clause call record; Collect the calculation basis for the damage failure probability and the judgment basis for the risk assessment conclusion, and form a conclusion support record; The reasoning path of the thought chain, the standard clause call record, and the conclusion support record are structurally associated according to the reasoning sequence to generate a traceable archive of the entire reasoning process; Based on the traceable archives, a natural language description of the risk assessment is output, including a summary of the reasoning chain, key judgment criteria, and risk assessment conclusions.

8. A risk assessment device for pressure equipment based on a knowledge graph-enhanced large language model, characterized in that, include: The reasoning tree generation module is configured to call the knowledge graph of pressure equipment, use each pressure equipment and its corresponding risk assessment association data as matching conditions, match and filter layer by layer in the multi-level damage and failure classification structure of the knowledge graph, extract the subset of damage and failure classification nodes that are suitable for the pressure equipment and have a tree topology, and generate the multi-level damage and failure reasoning tree corresponding to the pressure equipment. The node configuration module is configured to define the root node in the subset of damage and failure classification nodes as a private root node, and define its first intermediate branch node, second intermediate branch node, and leaf node as shared nodes that can be reused by all pressure-bearing equipment. Damage and failure knowledge rules are uniformly bound within all shared nodes. Independent association links are established between the private root node and each shared node, and each association link stores damage and failure constraint parameters that are only suitable for the corresponding pressure-bearing equipment operating conditions in an isolated manner. The hierarchical reasoning module is set up to be a thought chain reasoning path based on a tree topology hierarchy, forming a mandatory step-by-step reasoning constraint. This allows the large language model to perform hierarchical reasoning in the order of the tree topology hierarchy, outputting the damage failure category and its corresponding sub-damage mode of each branch node level by level. After retrieving all the required parameters by using the damage mode output by the leaf node as an index, the damage failure probability calculation model bound to the private root node performs a one-time summary calculation to obtain the damage failure probability. The risk level determination module is set to determine the risk level of the target pressure equipment based on the failure probability level corresponding to the damage failure probability and the failure consequence level corresponding to the damage failure constraint parameter, and generate a risk assessment conclusion. The traceable archive module is configured to archive the reasoning path of the thought chain, the industry standard clauses corresponding to each branch node, the calculation basis for the probability of damage and failure, and the judgment basis for the risk assessment conclusion, forming a traceable archive and outputting a natural description of the risk assessment of the pressure equipment.

9. An electronic device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 7, which, when invoked by a computer, executes the steps included in the corresponding method.

Images