Static equipment predictive maintenance large model agent construction method, device, equipment, medium and product

Abstract

The application discloses a large model intelligent agent construction method and device for predictive maintenance of static equipment, equipment, medium and product, relates to the field of industrial equipment predictive maintenance and artificial intelligence, and comprises the following steps: acquiring multi-source heterogeneous knowledge required by predictive maintenance in the whole life cycle operation of static equipment, constructing a static equipment predictive maintenance field knowledge base, obtaining modal feature vectors of multi-modal data in the knowledge base, pre-training the vectors based on a large language model as a base, giving the large model after pre-training the planning and execution ability of each static equipment predictive maintenance tool, obtaining a large model intelligent agent, the large model intelligent agent making a predictive decision on an automatic trigger condition or a to-be-processed predictive maintenance task instruction, generating a response decision result, outputting the response decision result and maintaining the static equipment according to the response decision result. The intelligent agent constructed by the application can make full use of large model technology, understand complex context, fuse multi-source knowledge, and make autonomous reasoning and decision.

Description

Technical Field

[0001] This application relates to the fields of predictive maintenance of industrial equipment and artificial intelligence, and in particular to a method, apparatus, equipment, medium and product for constructing a large-scale intelligent agent for predictive maintenance of static equipment. Background Technology

[0002] In process industries such as petrochemicals, power generation, and metallurgy, static equipment (such as pressure vessels, pressure pipelines, and storage tanks) in complete sets of equipment are the core infrastructure to ensure production continuity and safety. Their long-term, reliable, and stable operation is directly related to the economic benefits and safe production of the entire plant. Scientific, efficient, and economical maintenance strategies are the key to achieving long-term safe and stable operation of the plant.

[0003] Currently, the maintenance of static equipment generally adopts a preventive maintenance model based on fixed cycles, which mainly includes two levels: first, periodic inspections conducted according to safety technical specifications, with a cycle of 3 to 6 years; and second, annual inspections and routine maintenance organized by the company itself. This fixed-cycle maintenance strategy fails to fully consider the actual operating status of the equipment, the damage evolution process, and the dynamic changes in risks, which can easily lead to "over-maintenance," resulting in waste of resources and unnecessary production interruptions; or "under-maintenance," allowing potential hazards to develop into accidents.

[0004] To address these issues, risk-based inspection (RBI) technology has been widely adopted. RBI identifies potential damage modes in equipment, assesses the likelihood and consequences of failure, quantifies risk levels, and formulates differentiated inspection strategies accordingly, achieving a better balance between safety and cost-effectiveness. However, RBI is essentially a periodic management method focused on optimizing inspection plans. Its assessment results rely on manual input and historical data, and it has inherent limitations in real-time dynamic response, accurate prediction, and autonomous decision-making capabilities.

[0005] In recent years, technologies represented by generative artificial intelligence and large language models have demonstrated deep semantic understanding, complex reasoning, and content generation capabilities, providing a new paradigm for solving problems such as industrial system optimization, fault prediction, and decision-making. However, in the process industry, where professional barriers are high and safety requirements are stringent, the implementation of large model technology is still in its early stages, facing challenges such as difficulty in acquiring high-quality data and high requirements for knowledge integration and interpretability. Specifically, for predictive maintenance of static equipment, the technology currently faces three major bottlenecks: (1) The information silos reflecting the safety status of equipment are serious. Predictive maintenance involves complex data and multi-source heterogeneous data types (sensor data, historical records, drawings, documents, etc.). It is difficult to analyze and fuse multi-source heterogeneous data, extract features, and construct rule-based datasets. (2) Damage diagnosis, risk assessment, life prediction, and maintenance decisions are highly dependent on the experience of domain experts. This knowledge usually exists in unstructured documents or in the minds of experts and is difficult to be effectively extracted and utilized by traditional mathematical models. (3) Maintenance decisions need to take into account multiple constraints such as equipment status, process, resources, and plans. Traditional models or rule systems have weak generalization capabilities and are difficult to achieve adaptive intelligent decision-making.

[0006] In summary, current predictive maintenance adopts a periodic, passive maintenance approach, which lacks the technology to fully utilize large-scale modeling techniques and intelligent agents capable of understanding complex contexts, integrating multi-source knowledge, and performing autonomous reasoning and decision-making for predictive maintenance. This results in problems of under-maintenance or over-maintenance, and insufficient dynamic response in hazard identification and process risk mitigation measures. Summary of the Invention

[0007] The purpose of this application is to provide a method, apparatus, device, medium, and product for constructing a large-model intelligent agent for predictive maintenance of static equipment, which can fully utilize large-model technology and understand complex contexts, integrate multi-source knowledge, and perform autonomous reasoning and decision-making.

[0008] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for constructing a large-scale intelligent agent for predictive maintenance of static equipment, including: Acquire multi-source heterogeneous knowledge required for predictive maintenance of static equipment throughout its entire life cycle, and construct a domain knowledge base for predictive maintenance of static equipment.

[0009] Feature extraction and vector representation are performed on multimodal data in the knowledge base for predictive maintenance of static equipment to obtain modal feature vectors of multimodal data.

[0010] Using a large language model as a foundation, the large language model is pre-trained based on the modal feature vectors of modal data to complete the adaptive training of the large language model in the field of predictive maintenance, enabling the large language model to master the professional knowledge of static equipment and obtain the pre-trained large model.

[0011] By endowing the pre-trained large model with the ability to plan and execute predictive maintenance tools for various static equipment, a large model intelligent agent is obtained.

[0012] The large model agent makes predictive decisions on automatic triggering conditions or user-inputted predictive maintenance task instructions, and generates response decision results corresponding to the automatic triggering conditions or predictive maintenance task instructions; it outputs the generated response decision results and performs maintenance on static equipment based on the response decision results.

[0013] Secondly, this application provides a large-scale intelligent agent construction device for predictive maintenance of static equipment, comprising: The domain knowledge base construction and storage module is used to acquire multi-source heterogeneous knowledge required for predictive maintenance during the entire life cycle of static equipment, and to construct a domain knowledge base for predictive maintenance of static equipment.

[0014] The feature extraction module is used to extract features and represent vectors from multimodal data in the knowledge base for predictive maintenance of static equipment, and obtain modal feature vectors of the multimodal data.

[0015] The pre-training module is used to pre-train the large language model based on the modal feature vectors of modal data, thereby completing the adaptive training of the large language model in the field of predictive maintenance, enabling the large language model to master the professional knowledge of static equipment, and obtaining the pre-trained large model.

[0016] The agent construction module is used to endow the pre-trained large model with the ability to plan and execute predictive maintenance tools for various static equipment, thus obtaining the constructed large model agent.

[0017] The response decision generation and static equipment maintenance module is used by the large model agent to make predictive decisions on automatic triggering conditions or user-inputted predictive maintenance task instructions, generate response decision results corresponding to the automatic triggering conditions or predictive maintenance task instructions, output the generated response decision results, and perform static equipment maintenance based on the response decision results.

[0018] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for constructing a large model intelligent agent for predictive maintenance of static equipment.

[0019] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for constructing a large model intelligent agent for predictive maintenance of static equipment.

[0020] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for constructing a large model intelligent agent for predictive maintenance of static equipment.

[0021] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method, apparatus, device, medium, and product for constructing a large-scale intelligent agent for predictive maintenance of static equipment. It constructs a knowledge base for predictive maintenance of static equipment by leveraging the multi-source heterogeneous knowledge required for predictive maintenance throughout the entire lifecycle of the equipment. Feature extraction and vector representation are performed on the multi-modal data in this knowledge base to obtain modal feature vectors. This addresses the problems of severe information silos reflecting equipment safety status, complex data involved in predictive maintenance, multi-source heterogeneous data types, and difficulties in parsing and fusing multi-source heterogeneous data, extracting features, and constructing rule-based datasets.

[0022] By using a large language model as a foundation and pre-training it based on modal feature vectors from modal data, the adaptability training of the large language model in the field of predictive maintenance is completed. A pre-trained model that understands predictive maintenance expertise is constructed, and the pre-trained model is endowed with the ability to plan and invoke task tools. A large model agent is constructed, and response decision results are generated based on the large model agent. Based on the response decision results, static equipment is maintained, realizing intelligent diagnosis, prediction, and decision generation for predictive maintenance tasks of static equipment. This solves the problem that experiential knowledge usually exists in unstructured documents or in the minds of experts, which is difficult to be effectively extracted and utilized by traditional mathematical models, and that traditional models or rule systems have weak generalization ability and are difficult to achieve adaptive intelligent decision-making.

[0023] In summary, the intelligent agent constructed in this application fully utilizes large model technology and is able to understand complex contexts, integrate multi-source knowledge, and perform autonomous reasoning and decision-making. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating a method for constructing a large-scale intelligent agent for predictive maintenance of static equipment, provided in one embodiment of this application.

[0026] Figure 2This is a flowchart illustrating a method for constructing a large-scale intelligent agent for predictive maintenance of static equipment, provided as another embodiment of this application.

[0027] Figure 3 This is a functional architecture diagram of a large model intelligent agent construction system for predictive maintenance of static equipment, provided in another embodiment of this application.

[0028] Figure 4 This is a screenshot of the corrosion diagnosis report for the top of atmospheric pressure tower T101 provided in an embodiment of this application.

[0029] Figure 5 This is a screenshot of the remaining life diagnostic report for the P-302 depressurized residue oil pipeline provided in one embodiment of this application.

[0030] Figure 6 Risk matrix diagram of 578 device units in an embodiment of this application.

[0031] Figure 7 A risk assessment result diagram provided for an embodiment of this application.

[0032] Figure 8 A risk matrix diagram provided for one embodiment of this application.

[0033] Figure 9 This is an example diagram illustrating the time-series data change and control of acid value (TAN) provided in an embodiment of this application.

[0034] Figure 10 This is a schematic diagram of the functional modules of a large model intelligent agent construction device for predictive maintenance of static equipment, provided in an embodiment of this application.

[0035] Figure 11 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0036] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] In one exemplary embodiment, such as Figure 1 As shown, a method for constructing a large-scale intelligent agent for predictive maintenance of static equipment is provided, including: Step 201: Acquire multi-source heterogeneous knowledge required for predictive maintenance during the entire life cycle of static equipment, and construct a domain knowledge base for predictive maintenance of static equipment.

[0039] Step 202: Perform feature extraction and vector representation on the multimodal data in the knowledge base for predictive maintenance of static equipment to obtain the modal feature vectors of the multimodal data.

[0040] Step 203: Using the large language model as a base, pre-train the large language model based on the modal feature vectors of the modal data to complete the adaptation training of the large language model in the field of predictive maintenance, so that the large language model can master the professional knowledge of static equipment and obtain the pre-trained large model.

[0041] Step 204: Assign the pre-trained large model the ability to plan and execute predictive maintenance tools for each static device, thus obtaining the constructed large model agent.

[0042] Step 205: The large model agent makes predictive decisions on the automatic triggering conditions or the user-inputted predictive maintenance task instructions, and generates response decision results corresponding to the automatic triggering conditions or the predictive maintenance task instructions; outputs the generated response decision results, and performs maintenance on the static equipment according to the response decision results.

[0043] In practical applications, the multi-source heterogeneous knowledge required for predictive maintenance of static equipment throughout its entire life cycle includes: relevant laws and standards, design specifications, process operation procedures, equipment ledgers, failure mode libraries, historical maintenance reports, online monitoring system data, expert experience rules, and other multi-source heterogeneous knowledge related to predictive maintenance of static equipment. This knowledge base for predictive maintenance of static equipment in complete sets of equipment is constructed to provide professional knowledge support for the training and reasoning of intelligent agents.

[0044] In another exemplary embodiment of this application, the method for constructing a large model agent for predictive maintenance of static equipment further includes: interactive feedback results from experts to correct, verify, or evaluate the generated response decision results, and incremental learning of the large language model based on the interactive feedback results periodically.

[0045] In another exemplary embodiment of this application, the large model agent makes predictive decisions on automatically triggered conditions or user-inputted predictive maintenance task instructions, and generates response decision results corresponding to the automatically triggered conditions or user-input predictive maintenance task instructions, specifically as follows: The large model agent retrieves knowledge related to automatic triggering conditions or pending predictive maintenance task instructions from the knowledge base of predictive maintenance of static equipment, and generates modal feature vectors of knowledge related to automatic triggering conditions or pending predictive maintenance task instructions.

[0046] The modal feature vectors of knowledge related to automatic triggering conditions or pending predictive maintenance task instructions are input into the large model agent. The large model agent generates response decision results corresponding to automatic triggering conditions or pending predictive maintenance task instructions by calling the target tool. The target tool is a static equipment predictive maintenance tool related to automatic triggering conditions or pending predictive maintenance task instructions.

[0047] In practical applications, predictive maintenance tasks include: damage mode identification, corrosion thinning rate calculation, corrosion cracking sensitivity assessment, dynamic risk assessment, inspection strategy generation, remaining life assessment, equipment compliance check, and operational data anomaly check.

[0048] In another exemplary embodiment of this application, the output of the generated response decision result is specifically as follows: The generated response decisions are presented and displayed through the user interface in the form of structured reports, visual charts, or interactive suggestions. The generated response decisions are explained and their reasoning process and basis are traced.

[0049] In another exemplary embodiment of this application, the remaining life assessment tool is one or more of the following: a Paris formula correction algorithm based on fracture mechanics, a remaining wall thickness calculation algorithm based on corrosion rate, a Monte Carlo simulation algorithm based on reliability theory, a BP neural network algorithm based on machine learning, a support vector machine algorithm based on machine learning, and a recurrent neural network algorithm based on machine learning.

[0050] In practical applications, feature extraction and vector representation are performed on multimodal data in the knowledge base for predictive maintenance of static equipment to obtain modal feature vectors. Using a large language model as a foundation, the model is pre-trained based on these modal feature vectors to achieve adaptability training for predictive maintenance, enabling the large language model to acquire professional knowledge of static equipment. This results in a pre-trained large model, specifically: For any predictive maintenance task instruction, the multimodal data related to the predictive maintenance task instruction in the static equipment predictive maintenance domain knowledge base are extracted, and feature extraction and vector representation are performed to obtain the modal feature vector corresponding to the predictive maintenance task instruction.

[0051] Based on the modal feature vectors corresponding to each predictive maintenance task instruction, the large language model is pre-trained using the modal feature vectors of the modal data. This completes the adaptation training of the large language model in the predictive maintenance field, enabling the large language model to master the professional knowledge of static equipment, and thus obtaining the pre-trained large model.

[0052] In another exemplary embodiment of this application, the static equipment predictive maintenance tool includes: a damage pattern recognition tool, a corrosion thinning rate calculation tool, a corrosion cracking sensitivity discrimination tool, a risk dynamic assessment tool, an inspection strategy generation tool, a remaining life assessment tool, an equipment safety status assessment tool, and an operational data anomaly check and early warning tool.

[0053] This application also provides a specific embodiment that details the above-described method for constructing a large-scale intelligent agent for predictive maintenance of static equipment, specifically including: like Figure 2 As shown, this application provides a method for constructing a large-scale predictive maintenance intelligent agent for complete sets of equipment, including: Step 1: Constructing a domain knowledge base for predictive maintenance of static equipment, integrating multi-source heterogeneous knowledge such as regulations and standards, design specifications, process procedures, equipment ledgers, maintenance reports, and expert rules; Step 2: Cleaning, aligning, and performing feature engineering on the multimodal data of static equipment, generating three types of feature vectors: structured, textual, and visual, to support the construction of a pre-trained model; Step 3: Using a general-purpose large language model as the base, and utilizing the domain knowledge base and multimodal feature vectors for pre-training, constructing a domain knowledge adaptation model. Step 4: Based on the pre-trained model, endow it with task planning and tool invocation capabilities, integrate professional algorithm tools, and form a predictive maintenance large-scale intelligent agent that can autonomously execute tasks; Step 5: The intelligent agent receives instructions or triggers tasks, integrates working condition data, retrieves knowledge, and tool results, conducts multi-step causal reasoning, and outputs intelligent diagnosis, prediction, and maintenance decisions; Step 6: The output, interpretation, and human-machine collaborative optimization of decision results are carried out through continuous learning mechanisms to optimize the model and strategy, thereby achieving iterative improvement of system performance. This method specifically includes the following steps: S1: Construct a domain knowledge base for predictive maintenance of static equipment.

[0054] The constructed knowledge base for predictive maintenance of static equipment includes multi-source heterogeneous knowledge related to predictive maintenance of static equipment such as pressure vessels, pressure pipelines, and storage tanks, including laws and standards, design specifications, process operation procedures, equipment ledgers, failure mode libraries, historical maintenance reports, online monitoring system data, and expert experience rules.

[0055] Regulations and standards include: the Special Equipment Safety Law, the Regulations on Safety Supervision of Special Equipment, the TSG series of safety technical specifications, the GB / GB / T series of national standards, industry standards, and enterprise standards for pressure equipment, as well as the ASME BPVC series and API series of internationally recognized specifications.

[0056] Design specifications include: equipment design calculation sheets, strength verification reports, material selection specifications, corrosion allowance settings, design life definition, process package documents, PFD or PID drawings, assembly drawings, manufacturing acceptance standards, non-destructive testing requirements, test pressure and acceptance criteria, etc.

[0057] The process operation procedures include: unit start-up and shutdown procedures, normal operation control procedures, abnormal operating condition handling procedures, process parameter control indicators, corrosion prevention, agent injection, corrosion inhibition, water washing, condensation removal and other corrosion and damage prevention operation systems and execution records.

[0058] The equipment register includes: equipment tag number, name, category, specifications, manufacturer, serial number, commissioning date, main material, welding material, wall thickness, diameter, design or operating pressure and temperature, media name and composition, previous modifications, material upgrades, structural changes, monitoring point layout information, etc.

[0059] The failure mode library includes five major categories of damage modes: corrosion thinning, environmental cracking, material degradation, mechanical damage, and other damage. Each mode includes a mechanism description, sensitive materials or media or operating conditions, inducing conditions, morphological characteristics, detection methods, evolution laws, control and prevention measures, etc.

[0060] Historical maintenance reports include: periodic inspection reports, non-destructive testing reports, RBI assessment reports, wall thickness measurement data, defect records and assessment conclusions, maintenance plans and acceptance results, failure incident investigation reports, cause analysis and rectification records, etc.

[0061] Online monitoring system data includes DCS process timing data, such as pressure, temperature, and flow rate; LIMS system media flow data, such as sulfur content and chloride ion concentration; and online probe and thickness measurement data from the CMS corrosion monitoring system.

[0062] Expert experience rules include: rapid damage assessment rules for typical working conditions, corrosion rate correction coefficients, rules for the correlation between abnormal parameters and deterioration, early warning thresholds, inspection strategies, maintenance timing, monitoring frequency, monitoring focus, and other experience-based decision-making logic and engineering practices.

[0063] S2: Construction of multimodal datasets and generation of corresponding modal feature vectors.

[0064] For the multimodal data in S1, after data cleaning, alignment and feature engineering, a comprehensive static equipment status dataset covering historical archives and real-time operating conditions is constructed. Further feature extraction and vector representation processing are performed on the comprehensive static equipment status dataset to form the feature vectors of the corresponding modes.

[0065] The data in the domain knowledge base is divided into three types: structural chemical condition data, natural language descriptive text data, and image visual feature data. Structural chemical condition data includes: equipment number, equipment name, material type, material grade, medium name, medium content, design temperature, design pressure, operating temperature, operating pressure, commissioning date, nominal wall thickness, etc. Natural language descriptive text data includes: descriptive data in natural language from documents such as quality certificates, inspection history, maintenance work orders, and non-destructive testing reports. Image visual feature data includes: process flow diagrams, non-destructive testing images such as PT, MT, RT, and UT from documents such as regulations and standards, operating procedures, inspection history, and PID / PFD, as well as macroscopic images of equipment, metallographic structures, and scanning electron microscope images.

[0066] Furthermore, in the feature extraction and vector representation processing of structured chemical condition data, the data is divided into discrete parameters and continuous parameters. Continuous parameters do not have a natural numerical relationship between data points (e.g., equipment name, equipment number, material type, material grade, medium name). Continuous parameters take values ​​from any real number range and have clear mathematical dimensions and numerical meanings (e.g., temperature, pressure, flow rate). A unified encoding mapping is established for discrete parameters, converting textual category information into numerical identifiers to obtain discrete features, ensuring that similar data have similar semantic positions in the vector space. For continuous parameters, dimensionless processing is performed to eliminate dimensional differences, and normalization is performed by grouping by equipment type to obtain continuous features, making the values ​​of similar equipment comparable while preserving their relative differences. The processed discrete and continuous features are then fused and concatenated to form the structured chemical condition feature vector.

[0067] For natural language descriptive text data, the first step is to remove low-quality text, eliminating invalid content such as garbled characters, empty text, duplicate paragraphs, irrelevant and redundant information, and formatting errors. Simultaneously, duplicate sentences and highly similar semantics from multiple source documents are eliminated to ensure the standardization and purity of the input data. After low-quality text removal, the valid text undergoes preprocessing operations such as sentence segmentation, noise reduction, punctuation standardization, and terminology unification. Lexical segmentation is then performed according to semantic integrity, dividing continuous long texts into sequences of lexical units with independent semantic meaning. Based on this, a pre-trained language model is used to semantically encode the lexical sequences, extracting features such as technical terms, entity relationships, contextual logic, and impairment descriptions. These extracted multi-dimensional features are mapped to a unified semantic vector space to obtain the text semantic feature vector.

[0068] For image visual feature data, such as non-destructive testing images of equipment, corrosion morphology photographs, process flow diagrams, and equipment structure diagrams, data cleaning and standardization, image feature analysis and extraction, and visual feature vectorization generation are performed. First, data cleaning and standardization are carried out. A convolutional neural network is used to score the image sharpness, noise level, and brightness uniformity, filtering out low-quality and invalid images. Repairable images are enhanced, denoised, sized, and contrast-corrected, and the input format is standardized. Then, image feature analysis is performed. Optical character recognition and image segmentation are used to extract textual information such as equipment tag numbers, parameter annotations, and process symbols. Simultaneously, defect areas are segmented and located, and morphological, texture, and positional features of defects such as cracks, corrosion, and bulges are extracted. The analyzed visual units are encoded using a pre-trained visual model to obtain visual feature vectors.

[0069] S3: Building pre-trained large models with domain knowledge adaptation.

[0070] We select an open-source or commercial general-purpose large language model with powerful reasoning and code generation capabilities as the base, and use the multimodal feature vectors obtained from S2 as the input for training the large language model to train a pre-trained large model with domain knowledge adaptation.

[0071] Specifically, considering the model's open-source controllability, context window length, and inference cost, a large-scale base model was selected. Optionally, the base model could be an open-source version of DeepSeek or LlamA, which, while maintaining powerful code and logical reasoning capabilities, support fully private deployment. The specific training process for the model is described below: (1) Model input data: During the model training process, the input data consists of two types of content. The first is the structured chemical condition feature vector, text semantic feature vector and visual feature vector obtained after S2 feature processing. The second is the true value label data formed from on-site operation and maintenance records, actual test results and expert review conclusions, which covers the actual damage type of equipment, the on-site measured corrosion rate, the actual remaining life of equipment, the risk classification of equipment and the standard diagnostic conclusions that conform to industry standards, etc., to provide objective reference for model learning.

[0072] (2) Model Output Data: Various input data are fused by a multimodal adapter and then fed into the large-scale model on the base for computation. On the one hand, the model generates unified feature representation data that integrates heterogeneous information from multiple sources. On the other hand, it outputs predictive maintenance business results such as damage pattern recognition, corrosion thinning rate calculation, corrosion cracking sensitivity discrimination, risk dynamic assessment, inspection strategy generation, remaining life assessment, equipment safety status assessment, and abnormal operation data inspection and early warning. Each predicted content is compared with the pre-entered true value label one by one, and a quantitative loss index is calculated accordingly, which intuitively reflects the degree of error between the model prediction results and the actual working conditions.

[0073] (3) Parameter adjustment logic: The loss index obtained from the model calculation is used as the optimization benchmark. The backpropagation algorithm is used to solve the gradient values ​​corresponding to the network parameters layer by layer. The internal parameters of the model are continuously optimized according to the gradient change trend, and the error between the prediction result and the real label is continuously reduced, thereby improving the model's adaptability and inference accuracy in the field of static equipment prediction and maintenance.

[0074] (4) Model parameters that need to be adjusted: During the training phase, fix the inherent parameters of the main body of the base model, and optimize the full set of network parameters of the multimodal feature adapter, the local network parameters of the upper layer of the main model, and the task output parameters for various maintenance services. At the same time, dynamically adjust the learning rate, loss weight and other related training parameters based on the training effect of each round.

[0075] (5) Formation of a large-scale model adapted to the domain of predictive maintenance: Continuous iterative optimization until the loss of the validation set tends to be stable, and when the core business indicators such as damage pattern recognition, corrosion thinning rate calculation, corrosion cracking sensitivity discrimination, risk dynamic assessment, inspection strategy generation, remaining life assessment, equipment safety status assessment and operation data anomaly inspection and early warning meet the requirements of on-site engineering applications, the model training ends, all optimized parameters are solidified, and a pre-trained large-scale model adapted to the domain knowledge is obtained.

[0076] (6) Enhanced Retrieval Generation to Improve Model Accuracy: A hierarchical retrieval system is constructed: The first layer performs precise matching based on key metadata such as equipment tag number and damage type to quickly recall relevant technical documents; the second layer utilizes vectorized semantic retrieval to find similar scenarios from legal provisions and historical case databases. The retrieval system integrates a hybrid search strategy to reorder retrieval results according to the authority of the source (e.g., national standards are higher than enterprise specifications). The retrieved knowledge fragments are dynamically concatenated with the current task context of the large model to form accurate enhanced prompts and reduce the illusion of the large model's intelligent agent.

[0077] S4: Planning of predictive maintenance tools for static equipment and formation of large-scale predictive maintenance intelligent agents.

[0078] For the domain-adapted pre-trained model formed in step S3, the ability to plan specific tools for predictive maintenance tasks is given, thus forming a large predictive maintenance model agent.

[0079] Tools related to predictive maintenance tasks include, but are not limited to, damage pattern recognition, corrosion thinning rate calculation, corrosion cracking sensitivity assessment, dynamic risk assessment, inspection strategy generation, remaining life assessment, equipment safety status assessment, and operational data anomaly detection and early warning tools. By establishing the task planning capabilities of these tools, the intelligent agent, upon receiving natural language instructions, can decompose and plan tasks, and invoke relevant predictive maintenance tools for diagnostic decision-making.

[0080] Specifically, the damage mode identification tool is based on the GB / T 30579 standard. Damage mode types include corrosion thinning, stress corrosion cracking, material degradation, and mechanical damage. In the specific identification of damage modes, based on relevant material, medium, and process operation data of the equipment, identification is performed according to the "damage mode characteristic + rule matching" method given in GB / T 30579. As shown in GB / T 30579, corrosion thinning includes 29 types, environmental cracking 16 types, material degradation 16 types, mechanical damage 11 types, and other damage 10 types.

[0081] The corrosion thinning rate calculation tool, corrosion cracking susceptibility discrimination tool, risk dynamic assessment tool, and inspection strategy generation tool are implemented according to GB / T 26610.1-GB / T26610.5 standards. Based on the identified damage modes, the system incorporates programs for corrosion rate calculation, cracking susceptibility discrimination, failure probability calculation, and failure consequence calculation to obtain the aforementioned results. The determination of corrosion rate and cracking susceptibility involves two methods: one is to embed the algorithm logic included in GB / T26610.4 into the system; the other is to calculate the corrosion rate and cracking susceptibility for damage modes not included in the standard, based on relevant literature and retrieved information. Dynamic risk assessment, based on static data imported into the EAM system, combines dynamic data imported from the DCS, LIMS, and CMS systems via the system interface to achieve dynamic risk updates. Inspection strategy generation is based on the assessed equipment risk level, equipment type, defect type, and material type, and uses a built-in logical algorithm to provide inspection strategy information such as equipment inspection time, inspection type, and inspection method.

[0082] The remaining life assessment tool adaptively selects matching algorithms to calculate life based on equipment type, damage mode, and usage scenario characteristics. Supported algorithms include fracture mechanics algorithm, corrosion rate algorithm, reliability theory algorithm, and machine learning intelligent algorithm.

[0083] Among them, the fracture mechanics algorithm is for crack-type defects. It uses the Paris formula and the modified model to calculate the crack propagation rate and critical instability size, and derives the number of cycles and remaining life. For elastoplastic fracture, it uses J integral and COD theory to evaluate the remaining strength and back-calculate the safe operating cycle.

[0084] The corrosion rate algorithm is applicable to wall thickness reduction damage. It calculates the minimum allowable wall thickness based on the material's allowable stress and operating parameters, and combines the measured wall thickness with a reasonable corrosion rate to calculate the remaining life according to the GB / T 150 standard procedure.

[0085] The reliability theory algorithm uses Monte Carlo simulation, treating parameters such as defect size and corrosion rate as random variables. It obtains the lifetime probability distribution through sampling simulation and outputs the remaining lifetime at a specified confidence level.

[0086] Machine learning algorithms are designed for scenarios with complex mechanisms and multiple variables coupled together. They employ BP neural networks, support vector machines, and recurrent neural networks to adapt to multi-dimensional data, small sample, and time-series data scenarios, enabling the prediction of equipment performance degradation and remaining lifespan.

[0087] Equipment safety condition assessment tools, based on relevant technical specifications, conduct safety assessments of damage and defects that occur in equipment during manufacturing and operation, determining whether the equipment can continue to be used and whether it can operate until the next inspection cycle. For defects in the manufacturing and operation stages, GB / T 19624 is used to assess fracture, plasticity, and fatigue failure of planar defects (cracks, lack of fusion, incomplete penetration, ≥1mm undercut) and volumetric defects (pits, porosity, inclusions, <1mm undercut). For defects in in-service equipment, GB / T 35013 is used to assess the safety condition of corrosion thinning, hydrogen-induced cracking, dents, misalignment, fire damage, creep damage, and low-temperature brittle fracture.

[0088] The operational data anomaly detection and early warning tool uses a multi-dimensional threshold system to identify anomalies in the DCS temperature and pressure data and the media content data of the LIMS system connected to the intelligent agent system. This is based on historical operating data of the equipment, operating procedures, and parameter safety boundary requirements. (For example, it uses the safety operating boundary set according to the operating procedure, dynamically calculates the normal fluctuation range based on the statistical characteristics of historical data, and sets warning lines for key parameters based on the corrosion control requirements of materials and media.) When real-time data exceeds the threshold, the system automatically triggers a graded early warning, providing warnings for parameters that exceed the set safety boundaries.

[0089] S5: Intelligent diagnosis, prediction, and decision generation for predictive maintenance tasks.

[0090] After receiving the user's natural language instruction, the large model intelligent agent retrieves the relevant knowledge and multimodal data required by the instruction from the constructed predictive maintenance domain knowledge base, performs feature extraction and vector representation on the multimodal data, obtains the modal feature vector of the multimodal data, inputs the multimodal feature vector into the constructed large model intelligent agent, the large model intelligent agent calls the tool according to the content corresponding to the task instruction, the large model intelligent agent performs multi-step reasoning and causal analysis, and generates differentiated predictions as maintenance decision results according to the specific scenario and human-computer interaction instructions. For example: (1) The user inputs the pending predictive maintenance task instruction in the form of an interactive window dialogue to query the status information of static equipment, such as the material, medium, current operating temperature, risk level, damage mode category, remaining life and other parameters of the equipment. The query results can be further investigated to find out the specific reasons, the source of the data and the decision-making process; (2) The user inputs the pending predictive maintenance task instruction to generate a predictive maintenance report corresponding to the instruction. For example, to generate a damage mode recognition report for a specific device, the corresponding report is generated according to the built-in template.

[0091] S6: Decision output, feedback and system optimization.

[0092] The system presents decision-making results in the form of structured reports, visual charts, or interactive suggestions through a user interface, explaining the results and tracing the reasoning process and basis. It provides feedback channels, enabling domain experts to correct, verify, or evaluate the agent's judgments. This interactive feedback is stored as new experience in the knowledge base, and continuous learning mechanisms optimize the model and strategies, achieving iterative improvements in system performance.

[0093] Structured reports are designed to meet the requirements of equipment record management and regulatory compliance. During the generation of structured reports, the large model agent maps the JSON data generated by the system into Word / PDF templates according to the format requirements of relevant standards. Examples of structured reports include equipment periodic inspection reports, risk assessment reports, and safety status assessment reports. Structured reports meet regulatory traceability requirements and can be directly archived for future reference.

[0094] The visualized charts cater to the real-time scheduling needs of field operators, transforming complex decision-making data into an intuitive and easy-to-understand visual language. They support status display, trend analysis, and rapid root cause identification, and provide real-time alarm threshold alerts. The system offers a digital twin view of the equipment, supporting zooming and precise equipment positioning. Mouse hover or clicks allow users to view equipment tag number, material, medium, process parameters, damage mode, risk level, remaining lifespan, and inspection history, with color-coded indicators for equipment operating status.

[0095] The risk matrix diagram provides risk matrix diagrams for the entire plant and individual installations. In the plant's risk matrix diagram, different colors indicate the number of devices with different risk levels; for example, devices with low risk, medium risk, medium-high risk, and high risk are displayed in green, yellow, orange, and red, respectively. Clicking the further links on the risk matrix diagram for each type of risk device will display a list of the corresponding number of devices. The device list displays basic information about the devices, their failure probability, failure consequences, risk level, and remaining lifespan.

[0096] Trend Forecast Chart: This chart provides a trend forecast that combines historical data, current status, and future predictions. The horizontal axis of the trend forecast chart is the time axis, and the vertical axis is the key performance indicators (such as process parameters, media flow, wall thickness, corrosion rate, etc. in DCS, LIMS, and CMS systems). Upper and lower limit warning values ​​are set for each parameter on the chart. When the time-series data exceeds the warning value, the system will issue a warning to the operator to pay attention to the parameter and formulate measures to control the parameter to fall back to a safe range.

[0097] Interactive Dialogue Window: The interactive dialogue window is the core interface for natural language interaction between the intelligent agent and domain experts. It supports experts in asking questions or inputting commands in natural language, such as querying device status, requesting risk assessments, and asking for maintenance suggestions. The window automatically links to the current dialogue context and the device's historical interaction records to maintain dialogue continuity. Experts can ask follow-up questions such as "Why did you make this judgment?" or "Please provide a comparative analysis with data from the same period last year." The system accurately understands user needs through intent recognition and entity extraction technologies. The response includes text explanations and integrates visual charts, thumbnails, key data cards, hyperlinks, and other formats, supporting click-to-expand detailed reports.

[0098] Regarding the traceability and interpretability of the output results, the large-scale model agent records the entire inference, decision-making, and output chain, ensuring that the conclusions are verifiable through attribution analysis, evidence presentation, and knowledge tracing. For example, in reports such as remaining lifespan and risk assessment, the agent clearly provides the calculation basis and key influencing factors and highlights them, linking them to the original data; historical cases and legal provisions cited in the inference are all provided with directly accessible traceability links, making the decision-making process transparent and verifiable.

[0099] In terms of human-machine collaboration and feedback mechanisms, the predictive performance of the intelligent agent is continuously improved through multi-level closed-loop interaction. Within the interactive interface, domain experts can directly intervene in the system output through explicit feedback, including correcting damage diagnosis conclusions, adjusting prediction parameters such as corrosion rates, and evaluating and correcting reasoning steps to ensure that diagnoses and decisions align with the actual situation on-site. Simultaneously, the system automatically captures implicit feedback from expert operations through behavioral analysis. When experts reject system suggestions or take stricter measures, the intelligent agent automatically records decision deviations, analyzes, and optimizes risk thresholds and judgment logic. All feedback data, after anonymization, is used for iterative system optimization: firstly, corrected data and typical cases are stored in an experience database as a basis for subsequent decisions, and frequently confirmed content is transformed into explicit business rules; secondly, high-quality feedback samples are used to efficiently fine-tune and periodically update the parameters of the domain-wide model, improving reasoning accuracy and adaptability. The system thus forms an iterative process of "intervention-recording-optimization," continuously enhancing reliability and practicality through ongoing human-machine collaboration.

[0100] This application constructs an intelligent agent capable of autonomously handling complex maintenance decision-making tasks, effectively solving the problems of knowledge utilization, dynamic multi-constraint decision-making, and human-machine mutual trust in decision results. It realizes the application of large model technology in the vertical field of predictive maintenance of static equipment in complete sets of equipment, thereby improving the safety, economy and intelligence level of risk prevention and control of complete sets of equipment.

[0101] This application also provides a large-scale intelligent agent system for predictive maintenance of static equipment in complete sets of equipment, corresponding to the above method. Figure 3 The system's functional architecture diagram shows that it comprises seven modules: M1: Domain knowledge base construction and storage module; M2: Multimodal data processing and feature generation module; M3: Domain knowledge adaptation large model pre-training module; M4: Agent construction and task tool planning module; M5: Diagnosis, prediction, and decision push module; M6: Human-computer collaborative interaction and feedback module; and M7: System management and security module.

[0102] The system is generally divided into four layers: basic support layer (M1, M2), core intelligence layer (M3, M4), application interaction layer (M5, M6), and security control layer (M7).

[0103] The foundational support layer is responsible for the access, governance, and knowledge-based transformation of data across the entire domain. Module M1 collects knowledge data from fields such as regulations and standards, design specifications, process operation procedures, equipment ledgers, failure mode libraries, historical maintenance reports, online monitoring system data, and expert experience rules. Module M2 performs feature extraction, which is used for the training and construction of large-scale intelligent agents in the vertical domain by module M3.

[0104] The core intelligent layer is the decision-making center of the system. Based on multimodal feature vectors, a general large model is trained to form an M3 pre-trained large model, which deeply understands and masters the professional knowledge of predictive maintenance of static equipment, and at the same time endows it with the ability to plan specific tools for predictive maintenance-related tasks, forming the M4 predictive maintenance large model intelligent agent.

[0105] The application interaction layer realizes the output and display of decisions. M5 transforms specific instructions from operators into specific business actions (such as work orders and reports) through the decision-making of the intelligent agent. M6 provides a human-computer interaction interface, allowing experts to review and correct decisions and feed new experience back to the system, thereby achieving iterative improvement of the system's predictive performance.

[0106] The security control layer is the foundation of the system. It runs through all levels, providing security safeguards such as authentication, access control, data encryption, and operation auditing to ensure the reliable operation of the system in industrial environments.

[0107] The following details the specific functions of each module.

[0108] M1: Domain Knowledge Base Construction and Storage Module. This module serves as the knowledge base for the intelligent agent system, enabling the systematic collection and structured storage of multi-source heterogeneous knowledge in the field of predictive maintenance of static equipment. It stores, modifies, and updates multi-source heterogeneous knowledge related to predictive maintenance of static equipment, including regulations and standards, design specifications, process operation procedures, equipment ledgers, failure mode libraries, historical maintenance reports, online monitoring system data, and expert experience rules. This provides predictive and data support for task execution, model supervision and fine-tuning, and retrieval enhancement generation in the predictive maintenance vertical domain.

[0109] M2: Modal Data Processing and Feature Generation Module. It handles full-domain data access, governance, feature extraction, and vector generation, serving as a bridge between raw data and model training. It connects to the domain knowledge data collected from M1, categorizing the collected data into three types: structured condition data, natural language descriptive text data, and image visual feature data. It performs data cleaning, anomaly removal, temporal alignment, and standardization. For different data types, it conducts specialized feature extraction: structured data encoding normalization, text data semantic vectorization, and visual data feature parsing, forming feature vectors for the corresponding modalities. This provides standardized input for the pre-training of the large-scale domain model in M3, addressing the pain points of difficult multi-source heterogeneous data fusion and poor model interpretability.

[0110] M3: Domain Knowledge Adaptation Large Model Pre-training Module. This module focuses on the domain-specific adaptation of general-purpose large models. Based on a general-purpose large language model, it receives multimodal feature vectors output from M2 and performs domain-specific pre-training. Through learning from massive amounts of static equipment domain knowledge, the model deeply masters professional terminology, damage mechanisms, failure patterns, evaluation standards, and reporting specifications, establishing semantic understanding and knowledge association capabilities specific to industrial scenarios. After pre-training, the model overcomes generalization limitations and possesses a professional reasoning foundation in the field of predictive maintenance of static equipment. This provides model support for M4 to build intelligent agents with task execution capabilities, laying the foundation for professional intelligent decision-making.

[0111] M4: Agent construction and task tool planning module.

[0112] This module serves as the system's decision-making hub, responsible for integrating pre-trained models and specialized tools to construct a large-scale intelligent agent with autonomous execution capabilities. Relying on the domain-adapted large-scale model output by M3, it integrates capabilities such as task planning and tool invocation, and encapsulates eight types of specialized tools, including damage pattern recognition, corrosion rate calculation, risk assessment, and remaining life prediction, forming a tool library. The intelligent agent can autonomously decompose complex maintenance tasks, match corresponding tools, analyze execution results, and perform multi-step causal reasoning, realizing the processing from task instructions to diagnostic conclusions. It is the core execution unit for the system to achieve integrated prediction, diagnosis, and decision-making.

[0113] M5: Diagnosis, prediction, and decision push module.

[0114] This module is responsible for the formatted output and delivery of intelligent decision-making results, streamlining the process from intelligent reasoning to business execution. It receives the reasoning results from the M4 intelligent agent, combines them with equipment operating conditions, process constraints, and maintenance requirements, and generates standardized, executable business deliverables, including but not limited to damage pattern recognition reports, risk assessment reports, maintenance work orders, tiered early warning notifications, and inspection plan optimization suggestions. It supports customized output formats based on roles and scenarios, adapting to the needs of different users such as managers and maintenance personnel. Through standard API interfaces, it connects to systems such as EAM, ERP, and DCS to achieve the push of decision-making instructions and real-time delivery of early warnings, promoting the concrete implementation of intelligent decisions and supporting closed-loop management of equipment operation and maintenance.

[0115] M6: Human-Machine Collaborative Interaction and Feedback Module. This module serves as the window for human-machine interaction, enabling efficient collaboration between the intelligent agent and domain experts and maintenance personnel, as well as system iterative optimization. It provides functions such as natural language dialogue, visual dashboards, and decision tracing: the dialogue window supports multi-turn interactive queries of equipment status and interpretation of diagnostic conclusions; the visual dashboards intuitively display equipment risk matrices, time-series data trends, damage distribution, and other information; the decision tracing function allows for tracing the reasoning basis and linking legal cases, improving the interpretability of AI decisions. Simultaneously, a feedback channel is established to support experts in correcting diagnostic conclusions, supplementing empirical rules, and evaluating the rationality of decisions. Feedback data flows back to the knowledge base and drives model iteration, providing training data for continuous model optimization and achieving iterative improvements in system performance.

[0116] M7: System Management and Security Module. This module runs through all levels of the system and is the fundamental support module for ensuring the stable, compliant, and secure operation of the system in an industrial environment. Its core functions include authentication, access control, data security, operation auditing, and model version management: it employs a multi-dimensional authentication mechanism and controls system functions and data access permissions based on role-based hierarchical management; through transmission encryption, storage anonymization, and hierarchical data protection, it ensures the security of core enterprise process and equipment data; it records user operations, model calls, and data modification logs throughout the entire process, achieving traceability and accountability; it manages the entire lifecycle of domain models and knowledge base versions, supports version rollback and update control, comprehensively prevents security risks, and ensures the system is adaptable to harsh industrial environments and operates stably and reliably.

[0117] This application also provides a specific case study of applying the above method to predictive maintenance of an atmospheric and vacuum distillation unit in a certain enterprise. This unit is the main crude oil processing unit of the entire plant, with an annual processing capacity of 10 million tons. It involves key equipment such as towers (e.g., primary distillation towers, atmospheric towers, vacuum distillation towers), heat exchangers, furnaces, and storage tanks. The towers are divided into upper and lower units, and the heat exchangers are divided into shell-side and tube-side units for management. The unit includes 39 units of 27 pressure vessels (excluding heat exchangers), 238 evaluation units of 119 heat exchangers, and 301 pipelines, totaling 578 equipment units. These devices operate for extended periods in a high-temperature, high-pressure, and corrosive environment of sulfur-containing crude oil, making maintenance difficult and requiring high safety standards. Specific steps include: 1. Construction of a knowledge base in the field of atmospheric and vacuum distillation units.

[0118] First, by connecting with existing information systems such as EAM, DCS, LIMS, and ERP through an industrial gateway, a domain knowledge base is constructed. This base includes regulations and standards, design specifications, process operating procedures, equipment ledgers, failure mode libraries, historical maintenance reports, online monitoring system data, and expert experience rules related to predictive maintenance of static equipment in atmospheric and vacuum distillation units. EAM system: Access to ledger information for 578 static equipment units, such as equipment number, material grade, process design parameters, commissioning date, maintenance work order records, etc.

[0119] DCS system: Real-time access to process parameters of key equipment such as atmospheric pressure towers and vacuum pressure towers, as well as connected pipelines (such as vacuum pressure tower line 1 and atmospheric pressure tower line 2), such as temperature, pressure, feed rate, reflux ratio, etc., with a sampling frequency of 1 minute.

[0120] LIMS system: Inputs crude oil and distillate oil media analysis data, such as sulfur content, hydrogen sulfide content, acid value, chloride ion content, pH value, etc., and updates daily.

[0121] Corrosion monitoring system: It integrates online corrosion rate probe monitoring and electromagnetic ultrasonic wall thickness monitoring data, such as online corrosion probes installed at high-risk locations like the top of atmospheric pressure towers and the outlet of vacuum furnaces, as well as electromagnetic ultrasonic thickness measurement data of equipment walls, to collect corrosion rate and equipment wall thickness data in real time.

[0122] ERP document system: Import data such as PID drawings of the equipment, factory quality certificates of the equipment, historical inspection reports, and non-destructive testing reports.

[0123] Regulations and Standards: *Special Equipment Safety Law*, *Regulations on Safety Supervision of Special Equipment*, and related regulations and standards, such as TSG 21 *Safety Technical Specification for Stationary Pressure Vessels*, TSG31 *Safety Technical Specification for Industrial Pipelines*, GB / T 150 *Pressure Vessels*, GB / T 42783 *Integrity Management of Complete Sets of Equipment*, GB / T 30579 *Damage Mode Identification for Pressure Equipment*, GB / T 26610 *Guidelines for Risk-Based Inspection of Pressure Equipment*, GB / T 19624 *Safety Assessment of In-Service Defective Pressure Vessels*, GB / T 35013 *Compliance Evaluation of Pressure Equipment*, etc.; Foreign Standards and Technical Documents: Including ASME codes and API standards, such as the ASME BPVC series codes, API 510, API 570, API 571, API 580, API 584, API 970, etc.

[0124] Design specifications include process design packages, engineering design drawings and documents, and key equipment calculation sheets during the design phase; equipment installation and inspection records and instrumentation and control system as-built data during the construction and building phases; and process technology operation procedures, production operation records, and equipment maintenance and repair archives during the operation and maintenance phases.

[0125] Process Operation Procedures: "Aggressive and Vacuum Distillation Unit Process Operation Procedures" includes unit start-up and shutdown procedures, normal operation control procedures, abnormal operating condition handling procedures, process parameter control indicators, corrosion prevention, agent injection, corrosion inhibition, water washing, condensation removal and other corrosion and damage prevention operation systems and execution records.

[0126] Equipment register: The equipment tag number, name, category, specifications, manufacturer, serial number, commissioning date, main body material, welding material, wall thickness, diameter, design / operating pressure and temperature, and medium name and composition data of the 27 pressure vessels (excluding heat exchangers), 39 units, 119 heat exchangers, 238 evaluation units, and 301 pipelines of the unit, totaling 578 equipment units.

[0127] Failure Mode Library: Based on the characteristics of atmospheric and vacuum distillation units, a failure mode library for various damages to key equipment components is established. This library includes five major damage modes: corrosion thinning, environmental cracking, material degradation, mechanical damage, and other damage. Each mode is described in terms of its mechanism, sensitive materials / mediums / operating conditions, inducing conditions, morphological characteristics, detection methods, evolution patterns, and control and prevention measures. Examples include brine and hydrochloric acid corrosion in the desalination system, low-temperature dew point (HCl-H2S-H2O) corrosion in the three towers (primary distillation tower, atmospheric distillation tower, and vacuum distillation tower top), high-temperature sulfide naphthenic acid corrosion at the bottom of the three towers, creep in the heating furnace tubes, and ammonium chloride under-scale corrosion in the heat exchanger tube bundles. These examples include mechanism descriptions, sensitive operating conditions, and characteristic maps. Record all equipment failure cases that have occurred since the unit started operation, such as leakage of the tube bundle of the atmospheric pressure tower top air cooler (2018), thinning and perforation of the inlet pipeline of the vacuum residue pump (2020), and corrosion cracking of the atmospheric pressure tower top reflux tank (2022), and record the failure process, cause analysis, treatment measures and effects in detail.

[0128] Historical maintenance reports include periodic inspection reports, non-destructive testing reports, RBI assessment reports, wall thickness measurement data, defect records and assessment conclusions, maintenance plans and acceptance results, failure accident investigation reports, cause analysis and rectification records, etc. for 27 pressure vessels (excluding heat exchangers), 119 heat exchangers and 301 pipelines in the unit.

[0129] Online monitoring system data includes DCS process timing data, LIMS system media flow data, and CMS corrosion monitoring system online probe and thickness measurement data.

[0130] Expert experience rules: Collect the experience of senior experts with in-depth understanding of equipment operation and maintenance to form expert experience rules for predictive maintenance of equipment operation. For example, "When the temperature at the top of the atmospheric pressure tower is lower than the dew point temperature and the sulfur content is >1.5%, the corrosion rate should be corrected by a factor of 2" and "When the material of the vacuum residue pipeline is 20# steel, the wall thickness monitoring frequency should be increased to once a month".

[0131] 2. Modal dataset construction and corresponding modal feature vector generation.

[0132] The multimodal data in the knowledge base is cleaned, aligned, standardized, feature extracted, and vector generated to form three types of feature vectors: structured, textual, and visual.

[0133] In the construction of the multimodal dataset, the top unit of the atmospheric pressure tower T101 is selected as a typical device. The process of extracting and vectorizing features from its structural chemical condition data, text data, and image data to generate structural chemical condition feature vectors, text semantic feature vectors, and visual feature vectors is explained.

[0134] 2.1 Generation of structural chemical condition feature vectors.

[0135] The structured chemical condition data is divided into discrete parameters and continuous parameters, which are processed separately and then concatenated to generate the final vector.

[0136] (1) Generation of discrete parameters and vectors.

[0137] Discrete parameters include equipment tag number, equipment type, main material, and media type. They have no natural numerical relationship and are generated into discrete vectors using a globally unified coding mapping.

[0138] A unified coding dictionary was established for all static equipment in the entire atmospheric and vacuum distillation unit. The coding rules for some fields are explained below: Equipment tag number: The equipment is numbered according to the order in the ledger. For example, if T101 is the 30th equipment in the ledger, then the equipment tag number of T101 is "30".

[0139] Equipment type: Different equipment types are numbered separately, such as tower = 0, heat exchanger shell side = 1, heat exchanger tube side = 2, pipeline = 3, storage tank = 4, safety valve = 5, etc.

[0140] Material: Different material types are numbered, such as 20# steel=1, 16MnR=2, 316L=2, 15CrMo=3.

[0141] Medium: Different medium types are numbered, such as crude oil from the top of the tower = 0, residual oil = 1, and oil and gas from the top of the tower = 2.

[0142] For the atmospheric pressure tower T101, it is the 30th device in the equipment ledger. The equipment type is tower, the material is 16MnR, and the medium is crude oil. According to the established unified coding dictionary, the discrete feature vector corresponding to its equipment tag number, equipment type, material and medium is [30,0,1,0].

[0143] (2) Generation of continuous parameters and vectors.

[0144] Continuous parameters include design temperature, operating temperature, design pressure, operating pressure, tower wall thickness, crude oil sulfur content, acid value, and years of operation. To eliminate dimensions, preserve distribution, unify the global feature space, adapt to the statistical characteristics of industrial data, and ensure the model's generalization and reproducibility, Z-score standardization based on the global mean μ and standard deviation σ is used to calculate and generate continuous vectors. The Z-score standardization formula is shown below: .

[0145] in, x It is the original data value. It is the mean of the data. It is the standard deviation of the data. It is the standardized value.

[0146] Original continuous data for T101: Design temperature 390℃, operating temperature 110℃, design pressure 0.32MPa, operating pressure 0.15MPa, wall thickness 18mm, sulfur content 1.2%, acid value 0.6mg / g, and service life 17 years. The mean μ and standard deviation σ of all towers in the atmospheric and vacuum distillation unit were statistically analyzed. The normalized values ​​of the continuity parameters calculated using the above normalization formula are shown in Table 1.

[0147] Table 1 Normalized values ​​of continuity parameters

[0148] Therefore, the continuous eigenvector generated by the continuity parameter is [0.83, 0.67, 0.80, 0.67, 0.75, 0.67, 0].

[0149] The discrete vector and the continuous vector are concatenated and fused to obtain the structural chemical condition feature vector [30, 0, 1, 0, 0.83, 0.67, 0.67, 0.75, 0.67, 0].

[0150] 2.2 Generation of text semantic feature vectors.

[0151] The T101 test report for the atmospheric pressure tower was cleaned, segmented, and semantically encoded to generate a text semantic vector.

[0152] Original text: Slight pitting corrosion was found on the 5th tray at the top of the T101 atmospheric pressure tower. The pits were approximately 2 to 3 mm in diameter and about 1.5 mm deep. The temperature at the top of the tower was 110 degrees Celsius, and the sulfur content of the crude oil was 1.2%, indicating a risk of low-temperature dew point corrosion. It is recommended that the top area be inspected with ultrasonic thickness measurement during the next overhaul.

[0153] The above text was standardized with professional terminology, numerical values ​​and units, and simplified to a more colloquial style, resulting in a semantically clean and formatted standard text. The cleaned text reads: "Slight pitting corrosion was found on the 5th tray at the top of atmospheric pressure tower T101. The pits are approximately 2mm–3mm in diameter and 1.5mm deep. The tower top temperature is 110℃, and the crude oil sulfur content is 1.2%, indicating a risk of low-temperature dew point corrosion. It is recommended that ultrasonic thickness measurement be performed on the tower top area during the next maintenance."

[0154] Using word segmentation rules adapted to industrial text, the cleaned text is divided into semantically complete and clearly defined word sequence, accurately preserving key entity information: Atmospheric pressure tower / T101 / tower top / 5th tray / found / slight pitting / corrosion pit / diameter / approx. 2mm–3mm / depth / approx. 1.5mm / tower top temperature / 110℃ / crude oil sulfur content / 1.2% / present / low temperature dew point corrosion / risk / recommendation / next overhaul / tower top area / key / ultrasonic thickness measurement.

[0155] A pre-trained BERT model optimized for industrial text is selected. The segmented word sequence is input into the model, which captures contextual semantic relationships through a self-attention mechanism. It learns the deep semantic relationships between equipment information, defect types, operating conditions, risk causality, and maintenance suggestions, outputting a text semantic feature vector. An example is provided: the values ​​of the first 12 feature vectors. [-0.026,0.061,-0.108,0.092,0.210,-0.172,0.045,-0.095,0.136,-0.062,0.190,-0.037].

[0156] III. Visual Feature Vector Generation.

[0157] The RT film image data of the weld seam at the top of the T101 atmospheric pressure tower was preprocessed and convolutional feature extraction was used to generate a visual vector.

[0158] Original image: A suspected crack of about 15mm long and 3mm deep exists in the weld area, with the background being a grayscale texture of the metal substrate.

[0159] Image preprocessing: denoising, enhancement, size normalization, and pixel value standardization; Feature extraction: Inputting an industrial ResNet-50 model, deep visual features such as cracks, edges, and textures are extracted through convolutional layers, and a dense vector is output after pooling dimensionality reduction; Example: The first 12 feature vector values ​​are given below. [0.315, -0.211, 0.159, -0.076, 0.248, -0.134, 0.095, -0.182, 0.224, -0.056, 0.171, -0.111].

[0160] 3. Construction of pre-trained large models with domain knowledge adaptation.

[0161] The DeepSeek-V4 open-source large-scale model is selected as the foundation. This model possesses strong logical reasoning, code generation, long context understanding, and private deployment capabilities, making it suitable for the data security and complex reasoning needs of the industrial field. Domain-specific pre-training is conducted using structured chemical condition feature vectors, textual semantic feature vectors, and visual feature vectors generated in the S2 step as input. Through large-scale training with professional data from atmospheric and vacuum distillation units, the model deeply masters the unit's professional terminology, process flow, and equipment structure, understands various damage mechanisms, failure modes, sensitive operating conditions, and corrosion evolution patterns, and is familiar with relevant regulations, standards, diagnostic logic, and reporting specifications. Simultaneously, a three-layer retrieval enhancement generation system is constructed: the first layer performs precise metadata matching based on equipment tag number, damage type, and equipment type; the second layer recalls similar operating conditions, cases, and defect information through vector semantic retrieval; and the third layer performs authoritative sorting according to national standards, industry standards, enterprise standards, and publication time. This retrieval enhancement generation mechanism effectively reduces model illusions, improves diagnostic accuracy, traceability, and compliance, and ultimately outputs a pre-trained large-scale model suitable for the atmospheric and vacuum distillation unit field.

[0162] 4. Task tool planning and predictive maintenance of large model intelligent agent construction.

[0163] Based on a domain-adapted pre-trained large-scale model, a predictive maintenance large-scale model agent integrating an intelligent agent engine and a professional tool library is constructed, possessing full-link capabilities in task planning, tool invocation, inference decision-making, and reflective verification. The agent incorporates task planning tools, including a damage pattern recognition tool based on GB / T 30579, a corrosion thinning rate calculation tool, a corrosion cracking sensitivity discrimination tool, a risk dynamic assessment tool based on GB / T 26610, an inspection strategy generation tool, a multi-algorithm fusion remaining life assessment tool, an equipment compliance check tool, and an operational data anomaly early warning tool. The agent's capabilities cover natural language understanding, complex task decomposition, execution step planning, automatic tool invocation and parameter matching, tool result parsing, multi-step causal reasoning, evidence tracing, decision generation, and reflective verification. During task execution, the agent receives natural language instructions, automatically decomposes sub-tasks, matches corresponding professional tools, invokes and retrieves domain knowledge, integrates multi-source information to conduct reasoning, generates diagnostic reports, and completes interpretable decision tracing, realizing intelligent execution of the entire predictive maintenance process for static equipment.

[0164] 4. Intelligent diagnosis, prediction, and decision generation for predictive maintenance tasks.

[0165] The following two cases illustrate the process of intelligent agent diagnosis, prediction, and decision generation: early warning and diagnosis of low-temperature dew point corrosion at the top of atmospheric pressure tower and assessment of the remaining life of vacuum residue oil pipeline.

[0166] 4.1: Case 1: Early warning and diagnosis of low-temperature dew point corrosion at the top of an atmospheric pressure tower.

[0167] Triggering condition: On March 15, 2026, the system detected abnormal corrosion probe data at the top of the atmospheric pressure tower. The corrosion rate suddenly increased from 0.15 mm / year to 0.42 mm / year, exceeding the set warning threshold (0.3 mm / year).

[0168] Intelligent agent task planning: Invoke the corrosion rate calculation tool to accurately calculate based on the current operating conditions; retrieve relevant regulations and cases on atmospheric pressure tower top corrosion from the knowledge base; call the LIMS interface to query recent changes in crude oil sulfur content; and generate a diagnostic report through comprehensive analysis.

[0169] The tool's execution result is as follows.

[0170] The corrosion rate calculator returned the following: Based on the API 581 and GB / T26610.4 logic models, the theoretical corrosion rate under the current operating conditions (tower top temperature 110℃, material 16MnR, acid value 0.6mg / g, pressure 0.15MPa, sulfur content 1.8%) is 0.38mm / year, which is basically consistent with the measured value.

[0171] Knowledge base retrieval recall: Description of "low temperature dew point corrosion" in GB / T 30579: When the wall temperature is lower than the acid dew point, sulfides and water vapor condense to form sulfurous acid, leading to severe corrosion; a case of leakage of the air cooler at the top of the atmospheric pressure tower in 2022, caused by the failure to adjust the top injection agent in time after the sulfur content of crude oil increased to 2.0%.

[0172] LIMS query results: The sulfur content of crude oil will gradually increase from 1.2% to 1.8% starting from March 1, 2026.

[0173] The reasoning process of the intelligent agent is as follows.

[0174] Step 1: An abnormal corrosion rate (0.15→0.42mm / year) was detected, triggering a diagnostic.

[0175] Step 2: LIMS query revealed that the sulfur content increased synchronously (1.2%→1.8%), indicating a time correlation.

[0176] Step 3: Search the knowledge base to confirm that the sensitive operating condition for "low temperature dew point corrosion" is sulfur content > 1.5% and temperature below the dew point (current tower top temperature is 110℃, which is lower than the calculated dew point of 125℃).

[0177] Step 4: Based on comprehensive judgment, the current intensified corrosion is caused by low-temperature dew point corrosion due to the increased sulfur content in crude oil.

[0178] Step 5: It is predicted that under the current operating conditions, if the sulfur content remains at 1.8%, the remaining service life will be shortened from the originally predicted 6 years to 2.8 years.

[0179] Step 6: It is recommended to immediately take measures to inject neutralizing agent and to conduct ultrasonic thickness measurement spot checks on the top of the tower during the next maintenance (6 months later).

[0180] Example of generated decision report Figure 4 As shown.

[0181] 4.2: Case 2: Assessment of the remaining life of a vacuum residue pipeline.

[0182] Equipment background: Vacuum residue pipeline (pipeline number: P-302), material: 20# steel, specification: Φ219×8mm, design temperature: 300℃, design pressure: 1.6MPa, operating temperature: 280℃, medium: vacuum residue (sulfur content: 2.2%).

[0183] User instruction: "Please assess the remaining life of the P-302 vacuum residue pipeline and provide inspection recommendations."

[0184] Agent task planning: The system retrieves equipment records, design parameters, and historical inspection data; utilizes corrosion rate calculation tools based on a high-temperature sulfur corrosion model; searches for similar cases and regulatory requirements; and generates a comprehensive evaluation report.

[0185] Tool execution result: Equipment record: The minimum measured wall thickness during the most recent inspection (May 2025) was 6.2 mm (outer side of the elbow), the initial wall thickness was 8.0 mm, the thickness decreased by 1.8 mm in two years, and the average annual corrosion rate was 0.9 mm / year.

[0186] Corrosion rate calculation: Based on McConomy's formula, the theoretical corrosion rate at 280℃ and 2.2% sulfur content is 0.85 mm / year, which is consistent with the actual measurement.

[0187] Failure Case Database Search: 2020 leakage case of a No. 20 steel slag oil pipeline of a similar unit of a certain enterprise (perforation occurred after 8 years of operation when the wall thickness was reduced to 4.5mm).

[0188] The reasoning process of the intelligent agent is as follows.

[0189] Step 1: Obtain the current minimum wall thickness of 6.2mm, the initial wall thickness of 8.0mm, and the thickness already reduced by 1.8mm.

[0190] Step 2: Calculate the remaining corrosion allowance: initial wall thickness 6.2mm - minimum allowable wall thickness 4.5mm = 1.7mm. The minimum allowable wall thickness is calculated according to GB / T 20801.

[0191] Step 3: Corrosion rate 0.9 mm / year, remaining life = 1.7 mm ÷ 0.9 mm / year ≈ 1.9 years.

[0192] Step 4: The search revealed that the plant-wide overhaul plan is scheduled for 1.6 years ago.

[0193] Step 5: Remaining lifespan 1.9 years > overhaul cycle 1.6 years, it is recommended to monitor operation until overhaul, but monitoring should be intensified.

[0194] Step 6: The corrosion rate is high. It is recommended to conduct 100% X-ray inspection on the elbow area during the next maintenance and consider upgrading the material.

[0195] The generated remaining life assessment report, such as Figure 5 As shown.

[0196] 5. Decision output, feedback and system optimization.

[0197] 5.1: Visual dashboard display.

[0198] The system provides the following visual interface on the large screen in the central control room of the device.

[0199] (1) Overall risks of the equipment.

[0200] The risk assessment results of the 578 equipment units in the device are shown in the figure below. Figure 6 Among them, green indicates low risk (healthy): 197 units; yellow indicates medium risk (caution): 274 units; orange indicates medium-high risk (warning): 106 units; and red indicates high risk (serious): 1 unit.

[0201] (2) Risks of each specific piece of equipment.

[0202] Below the device risk matrix diagram, a list of all equipment in the system is displayed. For each device, key information for predictive maintenance is shown, such as equipment number, name, material, damage mechanism, corrosion rate, risk level, and details. Clicking on details reveals further specific information about the risk assessment for that equipment. The following example uses Chang'er Middle Pipeline 250-P-189-2.5A2 to illustrate the equipment risk assessment results and risk matrix. Figure 7 and Figure 8 As shown.

[0203] (3) Time series data change and parameter early warning chart.

[0204] It displays data on equipment corrosion rate, temperature and pressure in DCS, and media flow in LIMS system over the past year, and indicates the safety boundary control thresholds for relevant parameters. Figure 9 This is an example of the time-series data change and control of acid value (TAN).

[0205] 5.2: Interactive dialog window.

[0206] Here is an example of an engineer interacting with an intelligent agent.

[0207] Engineer: Is the corrosion problem on T-101 serious? Intelligent Agent: Currently, the corrosion rate at the top of tower T-101 is 0.42 mm / year, a 60% increase compared to last month, mainly due to the crude oil sulfur content rising to 1.8%. Under current operating conditions, the remaining service life is approximately 2.8 years, and the risk level has risen from medium risk to medium-high risk. It is recommended to increase the neutralizing agent injection to 8 ppm and to focus on inspecting the top of the tower during the next overhaul.

[0208] Engineer: Why does increased sulfur content lead to accelerated corrosion? Intelligent Agent: From the evaporation line outlet at the top of the tower to the air-cooled inlet at the top (or the inlet of the oil / gas and crude oil heat exchanger at the top), hydrogen chloride dew point corrosion frequently occurs. This occurs when water vapor in the oil / gas condenses into liquid droplets, and hydrogen chloride gas dissolves in the water, forming hydrochloric acid dew point corrosion. H2S originates from hydrogen sulfide in crude oil and decomposition products of sulfides in crude oil above 260℃. H2S exacerbates and promotes the corrosion of hydrochloric acid, thus forming a strong HCl-H2S-H2O cyclic corrosion. Case studies of similar units in the knowledge base from 2022 show that when the sulfur content increases from 1.0% to 1.8%, the corrosion rate increases by 2-3 times.

[0209] Engineer: What conditions are required if we switch to a corrosion-resistant material now? Intelligent Agent: The following conditions need to be prepared: (1) The material needs to be upgraded to 316L, and the spare parts procurement cycle is 45 days; (2) The entire plant needs to be shut down for at least 3 days for replacement; (3) Welding requires 4 welders with argon arc welding qualifications. Considering that the remaining service life is still 2.8 years, it is recommended to monitor and run the system until the overhaul in 2025 and handle it together, which is more cost-effective. Do you need me to generate a material upgrade plan for backup? Engineer: Okay, generate the backup plan.

[0210] Intelligent Agent: The "T-101 Tower Top Material Upgrade Backup Plan" has been generated, including a spare parts list, construction plan, and budget estimate (approximately RMB 1.2 million), and has been stored in the EAM system for future use.

[0211] 5.3: Expert Feedback and System Optimization.

[0212] Explicit Feedback Case: After reviewing the T-101 diagnostic report, equipment expert Zhang clicked "Correction" in the "Root Cause Analysis" section, adding "Increased Sulfur Content" to "Increased Sulfur Content and Insufficient Injection Dosage," and adding the annotation "The top injection pump has been frequently malfunctioning recently, and the actual injection volume is lower than the set value." The system saves this correction to the experience knowledge base and updates the rule: "When the corrosion rate is abnormal, in addition to the sulfur content, the operating status of the injection system must be checked simultaneously."

[0213] Implicit feedback capture: The system suggested that P-302 be "monitored until overhaul", but the equipment supervisor, Engineer Li, directly clicked "create emergency maintenance work order" and requested that the pipeline be replaced in advance. The system recorded this decision deviation and found that the pipeline was located on the critical path of the unit. Once leaked, it would cause the entire unit to shut down, and the severity of the consequences may have been underestimated. In subsequent decisions, the system automatically increased the consequence weight of critical path equipment.

[0214] The above embodiments fully demonstrate the practical application effect of the large-model intelligent agent construction method for predictive maintenance of static equipment provided in this application in the predictive maintenance of static equipment in a certain enterprise's atmospheric and vacuum distillation unit. By constructing a large-model intelligent agent enhanced with domain knowledge, the system realizes deep fusion of multi-source heterogeneous data, autonomous planning and execution of complex maintenance tasks, interpretable intelligent decision generation, and a continuously optimized human-machine collaboration mechanism. This effectively solves the problems of "under-maintenance" or "over-maintenance" in traditional maintenance models, providing a large-model intelligent agent solution for the intelligent maintenance of static equipment in the process industry.

[0215] This application constructs a large vertical domain model enhanced with domain knowledge, and uses this model as the core to drive an intelligent agent with perception, decision-making and execution capabilities. This fully utilizes the large model's capabilities in multimodal understanding, temporal reasoning, knowledge fusion and causal analysis to form a closed loop of "perception-diagnosis-prediction-decision", thereby realizing intelligent diagnosis of the health status of static equipment, accurate risk prediction and adaptive maintenance decision support.

[0216] This application has the following beneficial effects: 1. Deep Utilization and Explicitization of Knowledge. Traditional models struggle to leverage unstructured expert experience and massive amounts of document knowledge. This application achieves efficient utilization of tacit knowledge of static equipment, complex procedures, and historical cases by constructing a vectorized domain knowledge base and fine-tuning a large domain model. It can automatically complete end-to-end complex tasks from status assessment and risk ranking to generating specific maintenance strategies (timing, method, scope), reducing reliance on manual analysis.

[0217] 2. Personalized maintenance strategies with dynamic and adaptive decision-making. Unlike static plans based on fixed cycles or risks, the intelligent agent in this application can comprehensively consider multiple dynamic constraints such as real-time equipment status, process load, maintenance resources, and production plans to generate maintenance strategies based on status requirements, promoting the transformation from "on-time maintenance" to "on-demand maintenance".

[0218] 3. Trustworthy Human-Machine Collaboration and Continuous Evolution. A transparent and trustworthy human-machine collaboration interface is constructed through natural language interaction, visual tracing, and decision interpretation. Domain experts can guide, verify, or correct the agent's decisions. Simultaneously, the large-scale model agent, through memory and feedback mechanisms, can transform the results of each interaction and maintenance into new experiential knowledge, achieving continuous self-optimization and evolution of system performance.

[0219] Based on the same inventive concept, this application also provides a large-scale model agent construction device for predictive maintenance of static equipment, which implements the above-described method for constructing large-scale model agents for predictive maintenance of static equipment. The solution provided by this device is similar to the solution described in the above-described method. Therefore, the specific limitations of one or more embodiments of the large-scale model agent construction device for predictive maintenance of static equipment provided below can be found in the limitations of the large-scale model agent construction method for predictive maintenance of static equipment described above, and will not be repeated here.

[0220] In one exemplary embodiment, such as Figure 10 As shown, a large-scale intelligent agent construction device for predictive maintenance of static equipment is provided, comprising: The domain knowledge base construction and storage module is used to acquire multi-source heterogeneous knowledge required for predictive maintenance during the entire life cycle of static equipment, and to construct a domain knowledge base for predictive maintenance of static equipment.

[0221] The feature extraction module is used to extract features and represent vectors from multimodal data in the knowledge base for predictive maintenance of static equipment, and obtain modal feature vectors of the multimodal data.

[0222] The pre-training module is used to pre-train the large language model based on the modal feature vectors of modal data, thereby completing the adaptive training of the large language model in the field of predictive maintenance, enabling the large language model to master the professional knowledge of static equipment, and obtaining the pre-trained large model.

[0223] The agent construction module is used to endow the pre-trained large model with the ability to plan and execute predictive maintenance tools for various static equipment, thus obtaining the constructed large model agent.

[0224] The response decision generation and static equipment maintenance module is used by the large model agent to make predictive decisions on automatic triggering conditions or user-inputted predictive maintenance task instructions, generate response decision results corresponding to the automatic triggering conditions or predictive maintenance task instructions, output the generated response decision results, and perform static equipment maintenance based on the response decision results.

[0225] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 11 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data for building large-scale intelligent agents for predictive maintenance of static devices. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for building large-scale intelligent agents for predictive maintenance of static devices.

[0226] Those skilled in the art will understand that Figure 11 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.

[0227] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method embodiments.

[0228] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the above-described method embodiments.

[0229] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described method embodiments.

[0230] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0231] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0232] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, etc., and are not limited to these.

[0233] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0234] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for constructing a large-scale intelligent agent for predictive maintenance of static equipment, characterized in that, The method for constructing a large-scale intelligent agent for predictive maintenance of static equipment includes: Acquire multi-source heterogeneous knowledge required for predictive maintenance during the entire life cycle of static equipment, and construct a domain knowledge base for predictive maintenance of static equipment; Feature extraction and vector representation are performed on multimodal data in the knowledge base for predictive maintenance of static equipment to obtain modal feature vectors of multimodal data; Based on the large language model, the large language model is pre-trained using modal feature vectors from modal data to complete the adaptive training of the large language model in the field of predictive maintenance, enabling the large language model to master the professional knowledge of static equipment and obtain the pre-trained large model. By endowing the pre-trained large model with the ability to plan and execute predictive maintenance tools for various static equipment, a large model intelligent agent is obtained. The large model agent makes predictive decisions on automatic triggering conditions or user-inputted predictive maintenance task instructions, and generates response decision results corresponding to the automatic triggering conditions or predictive maintenance task instructions; it outputs the generated response decision results and performs maintenance on static equipment based on the response decision results.

2. The method for constructing a large-scale intelligent agent for predictive maintenance of static equipment according to claim 1, characterized in that, The method for constructing a large-scale intelligent agent for predictive maintenance of static equipment also includes: experts correcting, verifying, or evaluating the generated response decision results to obtain interactive feedback results, and periodically performing incremental learning on the large language model based on the interactive feedback results.

3. The method for constructing a large-scale intelligent agent for predictive maintenance of static equipment according to claim 1, characterized in that, The large model agent makes predictive decisions based on automatically triggered conditions or user-inputted predictive maintenance task instructions, generating response decision results corresponding to the automatically triggered conditions or user-input predictive maintenance task instructions, specifically: The large model agent retrieves knowledge related to automatic triggering conditions or pending predictive maintenance task instructions from the knowledge base of predictive maintenance of static equipment, and generates modal feature vectors of knowledge related to automatic triggering conditions or pending predictive maintenance task instructions. The modal feature vectors of knowledge related to automatic triggering conditions or pending predictive maintenance task instructions are input into the large model agent. The large model agent generates response decision results corresponding to automatic triggering conditions or pending predictive maintenance task instructions by calling the target tool. The target tool is a static equipment predictive maintenance tool related to automatic triggering conditions or pending predictive maintenance task instructions.

4. The method for constructing a large-scale intelligent agent for predictive maintenance of static equipment according to claim 1, characterized in that, The generated response decision results are output as follows: The generated response decisions are presented and displayed through the user interface in the form of structured reports, visual charts, or interactive suggestions. The generated response decisions are explained and their reasoning process and basis are traced.

5. The method for constructing a large-scale intelligent agent for predictive maintenance of static equipment according to claim 1, characterized in that, Predictive maintenance tools for static equipment include: damage mode identification tools, corrosion thinning rate calculation tools, corrosion cracking sensitivity discrimination tools, risk dynamic assessment tools, inspection strategy generation tools, remaining life assessment tools, equipment safety status assessment tools, and operational data anomaly detection and early warning tools.

6. The method for constructing a large-scale intelligent agent for predictive maintenance of static equipment according to claim 5, characterized in that, The remaining life assessment tool is one or more of the following: Paris formula correction algorithm based on fracture mechanics, remaining wall thickness calculation algorithm based on corrosion rate, Monte Carlo simulation algorithm based on reliability theory, BP neural network algorithm based on machine learning, support vector machine algorithm based on machine learning, and recurrent neural network algorithm based on machine learning.

7. A large-scale intelligent agent construction device for predictive maintenance of static equipment, characterized in that, The large-scale intelligent agent construction device for predictive maintenance of static equipment includes: The domain knowledge base construction and storage module is used to acquire multi-source heterogeneous knowledge required for predictive maintenance during the entire life cycle of static equipment, and to construct a domain knowledge base for predictive maintenance of static equipment. The feature extraction module is used to extract features and represent vectors from multimodal data in the knowledge base for predictive maintenance of static equipment, and obtain modal feature vectors of the multimodal data. The pre-training module is used to pre-train the large language model based on the modal feature vectors of modal data, using the large language model as a base. This completes the adaptive training of the large language model in the field of predictive maintenance, enabling the large language model to master the professional knowledge of static equipment and obtain the pre-trained large model. The agent building module is used to endow the pre-trained large model with the ability to plan and execute predictive maintenance tools for various static equipment, thus obtaining the constructed large model agent; The response decision generation and static equipment maintenance module is used by the large model agent to make predictive decisions on automatic triggering conditions or user-inputted predictive maintenance task instructions, generate response decision results corresponding to the automatic triggering conditions or predictive maintenance task instructions, output the generated response decision results, and perform static equipment maintenance based on the response decision results.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the large model agent construction method for predictive maintenance of static equipment as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for constructing a large model agent for predictive maintenance of static equipment as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the method for constructing a large model agent for predictive maintenance of static equipment as described in any one of claims 1-6.

Images