A dynamic diagnosis system and method for safety and reliability of a chemical plant

Abstract

The application discloses a chemical device production safety reliability dynamic diagnosis system, and belongs to the technical field of reliability evaluation, comprising a double-layer influence factor self-generation and weight dynamic updating function module, an influence factor knowledge graph module, a graph-enhanced space-time AI diagnosis module, and a safety reliability comprehensive evaluation and disposal feedback function module, which is used for performing source evidence, calculation caliber, data computability, barrier correlation and expert review and checking on candidate influence factors, generating current comprehensive weights of the influence factors according to the source evidence, graph transmission relationship, online diagnosis contribution and field feedback results, and identifying abnormal types, predicting risk levels, locating key contribution influence factors and root cause candidates in combination with a graph neural network and a GRU time sequence coding model. The system realizes dynamic diagnosis, interpretable evaluation, hierarchical early warning and closed-loop self-learning optimization of the safety reliability of chemical device production.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnosis of safety risks and evaluation of production safety and reliability in chemical plants, specifically to a dynamic diagnostic system and method for production safety and reliability in chemical plants. Background Technology

[0002] Chemical plants typically feature continuous operation, long process flows, closely interconnected equipment, strong parameter coupling, rapid anomaly propagation, and multiple layers of safety barriers. In actual production, the safety and reliability of these plants depend not only on whether process and equipment parameters such as temperature, pressure, flow rate, liquid level, valve position, and vibration exceed limits, but also on the combined influence of various factors, including equipment integrity, the operational status of safety instrumented systems, the timeliness of alarm responses, the closure of hazard rectification, the effectiveness of change management, personnel operational capabilities, emergency response capabilities, and external environmental disturbances. Furthermore, these factors are not independent but exhibit clear hierarchical relationships and transmission mechanisms. Therefore, the key to diagnosing the safety and reliability of chemical plants lies not only in identifying whether a single parameter is abnormal, but also in revealing the dynamic transmission mechanism of risk.

[0003] Existing methods for safety assessment, risk warning, and anomaly diagnosis in chemical plants still suffer from several shortcomings in engineering applications. These include unclear sources and formation mechanisms of safety and reliability influencing factors, insufficient expression of risk transmission relationships, a lack of explicit modeling of chemical plant processes, safety barriers, accident causal patterns, and management closed-loop status in AI anomaly detection models, and the absence of an effective closed loop between model training and on-site feedback. Therefore, there is an urgent need to propose a technical method for dynamic diagnosis of safety and reliability in chemical plant production. This method should organically combine multi-source influencing factor generation, plant-level knowledge graph construction, influencing factor weight updates, spatiotemporal AI diagnosis, comprehensive evaluation, and on-site feedback optimization to improve the accuracy, interpretability, and engineering application value of safety and reliability diagnosis in chemical plants. Summary of the Invention

[0004] The technical problem to be solved by this invention is to address the problems in the existing diagnosis of the safety and reliability of chemical plant production, such as unclear sources of influencing factors, insufficient basis for weight calculation, separation between AI diagnosis results and comprehensive evaluation results, and difficulty in updating the model through on-site feedback. The invention designs a dynamic diagnosis system and method for the safety and reliability of chemical plant production.

[0005] This invention proposes a dynamic diagnostic system for the safety and reliability of chemical plant production, comprising a dual-layer influencing factor self-generation and weight dynamic update module, an influencing factor knowledge graph module, a graph-enhanced spatiotemporal AI diagnostic module, and a comprehensive safety and reliability evaluation and handling feedback module. The system generates safety and reliability influencing factors through multi-source files and real-time plant data, generates the current comprehensive weight of the influencing factors through source evidence, knowledge graph relationships, online diagnostic contributions, and on-site feedback, and outputs a comprehensive safety and reliability score, risk level, key contribution influencing factors, root cause candidates, and handling suggestions based on the degree of influencing factor anomaly and the current comprehensive weight.

[0006] Preferably, the dual-layer impact factor self-generation and weight dynamic update function module includes a multi-source heterogeneous data fusion and cleaning sub-module. This sub-module is used to access enterprise production safety indicator databases, accident statistics results, production anomaly ledgers, business ledgers, expert experience, process cards, operating procedures, HAZOP / LOPA records, and data from DCS, SIS, GDS, PLC, MES, equipment status monitoring systems, and environmental monitoring systems. It performs field mapping, unit conversion, timestamp alignment, missing value imputation, outlier removal, and duplicate record merging on structured data. It performs text segmentation, entity recognition, risk object extraction, and source evidence location on unstructured text data. It performs sampling frequency unification, noise smoothing, steady-state window identification, threshold verification, and data quality labeling on real-time operating data, forming a standardized data foundation for impact factor generation, knowledge graph construction, and AI diagnosis.

[0007] Preferably, the dual-layer impact factor self-generation and weight dynamic update function module includes a general impact factor generation sub-module and a device-specific impact factor generation sub-module;

[0008] The general impact factor generation submodule is used to generate general impact factors applicable to different chemical plants based on the production safety index database of multiple large domestic chemical enterprises, accident statistics, and industry expert experience.

[0009] The device-specific influencing factor generation submodule is used to generate device-specific influencing factors that match the target device's operating mechanism and risk scenarios based on the target device's production anomaly data, enterprise anomaly ledger, business ledger, real-time operating data, as well as key factors such as corrosivity, negative pressure, dust, special material reactions, scaling and coking, equipment seal failure, and factors restricting the long-term operation of the device.

[0010] Preferably, the device-specific impact factor generation submodule employs a private retrieval enhancement generation method within the enterprise intranet and a locally deployed large language model to generate device-specific impact factors. The system performs local parsing, cleaning, segmentation, and vectorized indexing of device files, anomaly records, historical handling records, expert review records, and on-site observable data fields to form a private knowledge base containing text fragments, structured fields, device tag numbers, abnormal events, failure mechanisms, and safety barrier relationships. Furthermore, it utilizes industry safety terminology, accident cause descriptions, device anomaly cases, impact factor naming rules, and expert-annotated samples to train or efficiently fine-tune the locally deployed large language model. This enables the locally deployed large language model to generate candidate device-specific impact factors and their risk objects, associated tag numbers, failure mechanisms, corresponding safety barriers, abnormal risk states, calculation methods, statistical periods, evaluation directions, and source evidence chains based on the retrieved evidence fragments.

[0011] Preferably, the dual-layer impact factor self-generation and weight dynamic update function module further includes an impact factor entry gate control submodule. This submodule is used to perform synonym merging, name standardization, calculation standardization, data computability verification, barrier association verification, anomaly co-occurrence verification, and expert review on general impact factors and device-specific impact factors. For candidate impact factors with complete source evidence, field-observable fields, and correlation with safety barriers or abnormal risk states, and confirmed by experts, the system writes them into the core impact factor library. For candidate impact factors with risk significance but incomplete data collection conditions, the system writes them into the impact factor library to be built. For candidate impact factors with a single source, semantic repetition, or unclear calculation standards, the system writes them into the manual review library.

[0012] Preferably, the dual-layer influence factor self-generation and weight dynamic update function module further includes a source evidence weight calculation submodule, a graph relationship weight calculation submodule, an online diagnosis contribution weight calculation submodule, and a closed-loop feedback correction weight calculation submodule;

[0013] The source evidence weight calculation submodule is used to generate source evidence weights based on the frequency of enterprise adoption, the correlation strength of accident events, the correlation strength of production anomalies, data availability, and expert confirmation.

[0014] The graph relationship weight calculation submodule is used to generate graph relationship weights based on relationship type, process topology distance, anomaly co-occurrence frequency, safety barrier correlation degree, and risk propagation path confidence.

[0015] The online diagnostic contribution weight calculation submodule is used to generate online diagnostic contribution weights based on the state change magnitude, graph attention coefficient, anomaly score contribution, and the ranking results of key contribution factors.

[0016] The closed-loop feedback correction weight calculation submodule is used to correct the influence factor weights based on on-site confirmation, false alarm correction, treatment effect, recovery time, recurrence marker, and expert review results.

[0017] Preferably, the impact factor knowledge graph module organizes nodes according to the equipment and process topology layer, parameter and tag layer, impact factor layer, safety barrier layer, abnormal event and risk status layer, and safety and reliability evaluation layer, and establishes equipment connection relationship, parameter affiliation relationship, parameter to impact factor mapping relationship, impact factor to safety barrier representation relationship, safety barrier to abnormal status association relationship, and abnormal status to evaluation result transmission relationship.

[0018] Preferably, the graph-enhanced spatiotemporal AI diagnostic module takes the graph node feature matrix, adjacency matrix, relation weight matrix, and continuous time window sequence as input. It learns the spatial transmission relationships between devices, parameters, influencing factors, and security barriers through graph convolution, graph attention, or relational graph neural networks. Furthermore, it learns the dynamic patterns of influencing factor states changing over time through temporal convolutional networks, gated recurrent units, and GRUs. The graph-enhanced spatiotemporal AI diagnostic module first represents the influencing factor knowledge graph as follows: Where V is the set of nodes including equipment, tag parameters, influencing factors, safety barriers, abnormal risk states, and evaluation results; E is the set of relationship edges between nodes; R is the set of relationship types; and W is the relationship weight matrix. Within each time window, the system constructs a node feature matrix based on real-time operational data, alarm interlock data, equipment status data, and management ledger data. Construct an adjacency matrix based on graph relationships. Furthermore, the system constructs dynamic graph sequences from graph data within continuous time windows. During model training, it maps historical normal samples, abnormal event samples, handling and recovery samples, and expert review samples into dynamic graph sequences. Anomaly type, risk level, comprehensive score, key contribution factors, and root cause labels are used as supervision signals. The model aggregates spatial information through graph neural network layers.

[0019]

[0020] The evolution of node states within a continuous time window is learned through the GRU temporal coding layer:

[0021]

[0022] Finally, the diagnostic output layer outputs:

[0023]

[0024] in, This includes anomaly type, risk level, comprehensive safety and reliability score, key contribution factors, root cause candidates, and risk propagation paths. During online diagnosis, the system updates node features in real time according to a sliding time window and inputs them into the trained model to achieve safety and reliability status identification, short-term risk prediction, and interpretable diagnostic output.

[0025] Preferably, the comprehensive safety and reliability evaluation and handling feedback function module is used to output the operating status, anomaly type, risk level, comprehensive safety and reliability score, Top-K key contribution impact factors, risk propagation path, root cause candidate ranking, and handling suggestions. It also pushes the diagnostic results to the operator interface, where the operator confirms receipt, handling, and review status. Each anomaly diagnosis forms an event segment sample, which includes the anomaly start and end time, anomaly number, impact factor change, key contribution impact factor, root cause candidate, confirmed root cause, handling action, recovery time, false alarm cause, recurrence identifier, and review conclusion. The event segment samples are then fed back into the impact factor library, sample library, knowledge graph library, model parameter library, threshold library, and rule library. The comprehensive safety and reliability evaluation and handling feedback function module follows the... Calculate the overall security and reliability score for the current time window, where Let i be the current overall weight of the i-th influence factor. Let i be the degree of abnormality of the i-th influencing factor. This is a penalty item for safety barrier failure. This is a penalty item based on the severity of the abnormal event. Bonus points for handling and restoring the status quo.

[0026] The present invention also proposes a dynamic diagnostic method for the production safety and reliability of a chemical plant using the system described in any one of claims 1 to 9, comprising the following steps:

[0027] S1. Parse multi-source files and real-time data to form structured source units; generate candidate impact factors based on the source paths and perform data entry gating;

[0028] S2. Generate initial weights for impact factors based on source evidence; construct a knowledge graph of impact factors among equipment, parameters, impact factors, safety barriers, abnormal states, and evaluation results; generate a relation weight matrix based on the graph relationships;

[0029] S3. Input the graph node features, relation weight matrix and time window sequence into the graph-enhanced spatiotemporal AI diagnostic module;

[0030] S4. Generate dynamic contribution weights based on AI diagnostic results; integrate source evidence weights, graph relationship weights, dynamic contribution weights, and feedback correction weights to generate the current comprehensive weights;

[0031] S5. Calculate the comprehensive safety and reliability score based on the current comprehensive weight and the degree of abnormality of the impact factors;

[0032] S6. Feed the results of on-site confirmation, handling execution, restoration confirmation, false alarm correction and review back into the impact factor library, sample library, knowledge graph library, model parameters, threshold library and rule library.

[0033] Compared with existing technologies, the beneficial effects of this invention are: the impact factor can be automatically generated from multiple source paths and form an evidence chain; the impact factor weight can be jointly determined by source evidence, graph relationships, AI online contributions, and on-site feedback; the AI ​​diagnostic results can be interpreted along the path of "parameter—impact factor—safety barrier—abnormal risk state—comprehensive score"; and the on-site confirmation and handling review results can update the model and rules in reverse, thereby forming a traceable, interpretable, calculable, and closed-loop dynamic diagnostic system for the production safety and reliability of chemical plants. Attached Figure Description

[0034] Figure 1 This invention

[0035] Figure 2 Homepage image of the dynamic diagnostic system for production safety and reliability of chemical plants

[0036] Figure 3 Interface diagram of the two-layer impact factor self-generation and weight dynamic update function module

[0037] Figure 4 Impact Factor Knowledge Graph Module Interface Diagram

[0038] Figure 5 Schematic diagram of the graph-enhanced spatiotemporal AI diagnostic module

[0039] Figure 6 Interface diagram of the comprehensive safety and reliability evaluation and handling feedback function module Detailed Implementation

[0040] The present invention will be further explained below with reference to specific embodiments.

[0041] like Figure 1As shown, a dynamic diagnostic system for the safety and reliability of a chemical plant mainly includes: a dual-layer influencing factor self-generation and weight dynamic update module, an influencing factor knowledge graph module, a graph-enhanced spatiotemporal AI diagnostic module, and a comprehensive safety and reliability evaluation and response feedback module. This system relies on technologies such as multi-source data analysis, knowledge graphs, graph neural networks, temporal deep learning, and visual interaction to uniformly access and integrate chemical plant equipment operation data, alarm interlock data, equipment health data, hazard rectification data, personnel management data, accident statistics, production anomaly data, and industry expert experience. This enables dynamic identification of the safety and reliability status of the chemical plant, risk level prediction, key contributing influencing factor location, root cause path explanation, and response feedback updates. Figure 2 As shown, the system is applied to the atmospheric and vacuum distillation unit of a refining and chemical enterprise, and the main functional modules and front-end interface of the system are demonstrated.

[0042] The system can be developed in a Windows 11 environment and deployed using a B / S architecture and a relational database. The backend uses Python, Java, and other programming languages ​​to implement data parsing, model training, knowledge graph construction, graph neural network calculation, and scoring calculation functions. The frontend uses HTML5, CSS3, JavaScript, and other technologies to display the system interface. The database uses SQL Server to store the impact factor library, weight library, knowledge graph library, historical sample library, model output result library, and feedback record library. The system frontend interface includes pages such as homepage overview, impact factor library, weights and knowledge graph, AI diagnostic analysis, safety and reliability scoring, and handling suggestions and feedback updates.

[0043] Depend on Figure 3 As shown, the dual-layer impact factor self-generation and weight dynamic update function module operates according to the path of "multi-source heterogeneous data fusion and cleaning - generation of general impact factors + device-specific impact factors - data entry gating - weight dynamic update".

[0044] The system first integrates and cleans multi-source heterogeneous data, unifying and accessing enterprise performance indicator databases, accident statistics, production anomaly ledgers, business ledgers, expert experience, process cards, operating procedures, HAZOP / LOPA records, and data from DCS, SIS, GDS, equipment status monitoring, and environmental monitoring. Data is processed separately for structured data, unstructured text data, and real-time operational data. For structured data, the system performs field mapping, unit conversion, timestamp alignment, missing value imputation, outlier removal, and duplicate record merging. For text data, the system performs text segmentation, entity recognition, risk object extraction, and source evidence location. For real-time operational data, the system performs sampling frequency unification, noise smoothing, steady-state window identification, threshold verification, and data quality labeling. The cleaned data is then integrated according to unit, equipment tag number, data field, time window, risk event, and safety barrier to form a standardized data foundation.

[0045] The system generates general impact factors based on a database of production safety indicators from over 20 large domestic chemical enterprises, accident statistics, and industry expert experience. This general impact factor library is then formed through synonym merging, standardization, and manual review. Simultaneously, based on production anomaly data from the target unit, enterprise anomaly logs, real-time operational data, and operational technical guidelines for key factors restricting the unit's long-term operation, the system locally parses, cleans, segments, and vectorizes the unit files, anomaly records, historical handling records, expert review records, and observable field data fields on the enterprise's intranet. This forms a private knowledge base composed of text fragments, structured fields, unit tag numbers, anomaly events, failure mechanisms, and safety barrier relationships. Subsequently, the system efficiently fine-tunes the parameters of the local large-scale language model using industry safety terminology, accident cause descriptions, unit anomaly cases, impact factor naming rules, and expert-annotated samples. This enables the model to retrieve evidence fragments related to the target unit, anomaly scenarios, and operational constraints, and input the retrieval results and unit tag numbers into the local large-scale model to generate unit-specific impact factors. The system further verifies the computability of general impact factors and device-specific impact factors through data computability verification, barrier association verification, anomaly co-occurrence verification, and expert review. Based on the verification results, the data is written into the core impact factor library, the impact factor library to be built, or the manual review library. After the impact factors are entered into the library, the system generates the current comprehensive weight based on source evidence, graph transmission relationships, online diagnostic contributions, and field feedback results. This weight is then used for comprehensive safety and reliability evaluation, knowledge graph node and edge weight initialization, and message propagation constraints of the graph-enhanced spatiotemporal AI diagnostic model.

[0046] In one specific calculation method, the source evidence weight of the i-th impact factor can be expressed as:

[0047]

[0048] in, This represents the base weight of the i-th influence factor. Indicates the frequency of use by enterprises. Indicates the strength of the correlation between the accidents. Indicates the strength of the correlation between production anomalies. Indicates the spectral transmission intensity. Indicates the degree of expert confirmation. This is the adjustment coefficient.

[0049] The combined weight of the i-th influencing factor at time t can be expressed as:

[0050]

[0051] in, This represents the combined weight of the i-th influencing factor at time t. This indicates the online diagnostic contribution of the impact factor within the current time window. This indicates the feedback and correction results from the site. , , This is the fusion coefficient.

[0052] The normalized comprehensive weights are used for comprehensive safety and reliability scoring, knowledge graph node importance assignment, relation edge weight initialization, and spatiotemporal graph neural network message propagation constraints, enabling the model to prioritize key influencing factors that are reliable in origin, strongly correlated with accidents, sensitive to anomalies, have clear graph transmission paths, and have been verified to be effective by field feedback.

[0053] like Figure 4 As shown, the impact factor knowledge graph module employs a dynamic update mechanism that updates node status in real time and expands the graph structure as needed. The system updates the status characteristics of the tag parameter nodes and impact factor nodes based on real-time running data. When a new impact factor is generated, the system determines whether it already exists in the knowledge graph. If it does, the system updates the node attributes and weights; if it does not exist but meets the data source and calculation criteria requirements, a new impact factor node is added, and relationship edges are established between it and the tag parameter, safety barriers, abnormal risk states, and evaluation results. Through this mechanism, the knowledge graph can be dynamically updated according to changes in the device's operating status and the impact factor database.

[0054] In one specific implementation, such as Figure 5 As shown, the graph-enhanced spatiotemporal AI diagnostic module first represents the impact factor knowledge graph as follows: Where V is the set of nodes including equipment, tag parameters, influencing factors, safety barriers, abnormal risk states, and evaluation results; E is the set of relationship edges between nodes; R is the set of relationship types; and W is the relationship weight matrix. Within each time window, the system constructs a node feature matrix based on real-time operational data, alarm interlock data, equipment status data, and management ledger data. Construct an adjacency matrix based on graph relationships. The system constructs dynamic graph sequences from graph data within continuous time windows. During model training, historical normal samples, abnormal event samples, handling and recovery samples, and expert review samples are mapped to dynamic graph sequences, using anomaly type, risk level, comprehensive score, key contribution impact factors, and root cause labels as supervision signals. The model aggregates spatial information through graph neural network layers.

[0055]

[0056] The evolution of node states within a continuous time window is learned through the GRU temporal coding layer:

[0057]

[0058] Finally, the diagnostic output layer outputs:

[0059]

[0060] in, This includes anomaly type, risk level, comprehensive safety and reliability score, key contribution factors, root cause candidates, and risk propagation paths. During online diagnosis, the system updates node features in real time according to a sliding time window and inputs them into the trained model to achieve safety and reliability status identification, short-term risk prediction, and interpretable diagnostic output.

[0061] like Figure 6 As shown, the comprehensive safety and reliability evaluation and response feedback module is used to calculate the comprehensive safety and reliability score of the chemical plant within the current time window. Simultaneously, it pushes the anomaly type, risk level, comprehensive score, key contribution factors, risk propagation path, and response suggestions to the operator interface. The system first standardizes and unifies the direction of each influencing factor to obtain the degree of anomaly of the (i)th influencing factor at time (t). It also calls the current comprehensive weight output by the influence factor weight self-generation and graph coupling function module. The system then calculates the overall score using the following formula:

[0062]

[0063] in, This represents the overall security and reliability score for the current time window. This represents the contribution of the i-th influencing factor to the decrease in score. This indicates that the security barrier weakens the penalty. This indicates a penalty item for abnormal events or risk states. This indicates corrections made after on-site handling, restoration confirmation, or expert review. The system classifies risk levels based on a comprehensive score and follows... The system outputs the Top-K key contribution impact factors from largest to smallest; simultaneously, it uses an impact factor knowledge graph to trace the transmission path and outputs root cause candidate rankings and treatment recommendations. After receiving the information, operators can confirm the diagnostic results, root cause candidates, and treatment recommendations, and record the treatment actions, adjustment magnitudes, recovery times, and review conclusions. If confirmed as a genuine anomaly, the system increases the weight of relevant impact factors and propagation paths; if confirmed as a false alarm, the system decreases the confidence level of relevant place values, thresholds, or rules; if there is a short-term recurrence after treatment, the system increases the weight of the corresponding root cause and treatment failure factors.

[0064] Meanwhile, when the overall safety and reliability score is below 75, the system triggers a Level 1 warning and pushes the warning information to relevant functional departments; when the overall safety and reliability score is below 60, the system triggers a Level 2 warning and pushes a high-level warning information to company leaders and relevant management departments.

[0065] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic diagnostic system for the production safety and reliability of a chemical plant, characterized in that: The system includes a two-layer impact factor self-generation and dynamic weight update module, an impact factor knowledge graph module, a graph-enhanced spatiotemporal AI diagnosis module, and a comprehensive safety and reliability evaluation and handling feedback module. The system generates safety and reliability impact factors through multi-source files and real-time device data, generates the current comprehensive weight of the impact factors through source evidence, knowledge graph relationships, online diagnostic contributions, and on-site feedback, and outputs a comprehensive safety and reliability score, risk level, key contribution impact factors, root cause candidates, and handling suggestions based on the degree of impact factor anomaly and the current comprehensive weight.

2. The dynamic diagnostic system for the production safety and reliability of a chemical plant according to claim 1, characterized in that: The dual-layer impact factor self-generation and weight dynamic update function module includes a multi-source heterogeneous data fusion and cleaning sub-module. This sub-module is used to access enterprise production safety indicator databases, accident statistics, production anomaly ledgers, business ledgers, expert experience, process cards, operating procedures, HAZOP / LOPA records, and data from DCS, SIS, GDS, PLC, MES, equipment status monitoring systems, and environmental monitoring systems. It performs field mapping, unit conversion, timestamp alignment, missing value imputation, outlier removal, and duplicate record merging on structured data. It performs text segmentation, entity recognition, risk object extraction, and source evidence location on unstructured text data. It performs sampling frequency unification, noise smoothing, steady-state window identification, threshold verification, and data quality labeling on real-time operating data, forming a standardized data foundation for impact factor generation, knowledge graph construction, and AI diagnosis.

3. The dynamic diagnostic system for the production safety and reliability of a chemical plant according to claim 1, characterized in that: The dual-layer impact factor self-generation and weight dynamic update function module includes a general impact factor generation sub-module and a device-specific impact factor generation sub-module. The general impact factor generation submodule is used to generate general impact factors applicable to different chemical plants based on the production safety index database of multiple large domestic chemical enterprises, accident statistics, and industry expert experience. The device-specific influencing factor generation submodule is used to generate device-specific influencing factors that match the target device's operating mechanism and risk scenarios based on the target device's production anomaly data, enterprise anomaly ledger, business ledger, real-time operating data, as well as key factors such as corrosivity, negative pressure, dust, special material reactions, scaling and coking, equipment seal failure, and factors restricting the long-term operation of the device.

4. The dynamic diagnostic system for the production safety and reliability of a chemical plant according to claim 3, characterized in that: The device-specific impact factor generation submodule employs a private intranet retrieval enhancement generation method and a locally deployed large language model to generate device-specific impact factors. The system performs local parsing, cleaning, segmentation, and vectorized indexing of device files, anomaly records, historical handling records, expert review records, and on-site observable data fields to form a private knowledge base containing text fragments, structured fields, device tag numbers, abnormal events, failure mechanisms, and safety barrier relationships. It also utilizes industry safety terminology, accident cause descriptions, device anomaly cases, impact factor naming rules, and expert-annotated samples to train or efficiently fine-tune the locally deployed large language model. This enables the locally deployed large language model to generate candidate device-specific impact factors and their risk objects, associated tag numbers, failure mechanisms, corresponding safety barriers, abnormal risk states, calculation methods, statistical periods, evaluation directions, and source evidence chains based on retrieved evidence fragments.

5. The dynamic diagnostic system for the production safety and reliability of a chemical plant according to claim 1, characterized in that: The dual-layer impact factor self-generation and weight dynamic update function module also includes an impact factor database entry gate control submodule. This submodule is used to perform synonym merging, name standardization, calculation standardization, data computability verification, barrier association verification, anomaly co-occurrence verification, and expert review on general impact factors and device-specific impact factors. For candidate impact factors with complete source evidence, field-observable fields, and the ability to be associated with safety barriers or abnormal risk states, and which have been confirmed by experts, the system writes them into the core impact factor database. For candidate impact factors with risk significance but incomplete data collection conditions, the system writes them into the impact factor database to be built. For candidate impact factors with a single source, semantic repetition, or unclear calculation standards, the system writes them into the manual review database.

6. The dynamic diagnostic system for the production safety and reliability of a chemical plant according to claim 1, characterized in that: The dual-layer impact factor self-generation and weight dynamic update function module also includes a source evidence weight calculation submodule, a graph relationship weight calculation submodule, an online diagnosis contribution weight calculation submodule, and a closed-loop feedback correction weight calculation submodule; The source evidence weight calculation submodule is used to generate source evidence weights based on the frequency of enterprise adoption, the correlation strength of accident events, the correlation strength of production anomalies, data availability, and expert confirmation. The graph relationship weight calculation submodule is used to generate graph relationship weights based on relationship type, process topology distance, anomaly co-occurrence frequency, safety barrier correlation degree, and risk propagation path confidence. The online diagnostic contribution weight calculation submodule is used to generate online diagnostic contribution weights based on the state change magnitude, graph attention coefficient, anomaly score contribution, and the ranking results of key contribution factors. The closed-loop feedback correction weight calculation submodule is used to correct the influence factor weights based on on-site confirmation, false alarm correction, treatment effect, recovery time, recurrence marker, and expert review results.

7. The dynamic diagnostic system for the production safety and reliability of a chemical plant according to claim 1, characterized in that: The impact factor knowledge graph module organizes nodes according to the equipment and process topology layer, parameter and tag layer, impact factor layer, safety barrier layer, abnormal event and risk status layer, and safety and reliability evaluation layer. It also establishes equipment connection relationships, parameter affiliation relationships, parameter-to-impact factor mapping relationships, impact factor-to-safety barrier representation relationships, safety barrier-to-abnormal status association relationships, and abnormal status-to-evaluation result transmission relationships.

8. The dynamic diagnostic system for the production safety and reliability of a chemical plant according to claim 1, characterized in that: The graph-enhanced spatiotemporal AI diagnostic module takes the graph node feature matrix, adjacency matrix, relation weight matrix, and continuous time window sequence as input. It learns the spatial transmission relationships between devices, parameters, influencing factors, and security barriers through graph convolution, graph attention, or relational graph neural networks. Furthermore, it learns the dynamic patterns of influencing factor states over time through temporal convolutional networks, gated recurrent units, and GRUs. The graph-enhanced spatiotemporal AI diagnostic module first represents the influencing factor knowledge graph as follows: Where V is the set of nodes including equipment, tag parameters, influencing factors, safety barriers, abnormal risk states, and evaluation results; E is the set of relationship edges between nodes; R is the set of relationship types; and W is the relationship weight matrix. Within each time window, the system constructs a node feature matrix based on real-time operational data, alarm interlock data, equipment status data, and management ledger data. Construct an adjacency matrix based on graph relationships. Furthermore, the system constructs dynamic graph sequences from graph data within continuous time windows. During model training, it maps historical normal samples, abnormal event samples, handling and recovery samples, and expert review samples into dynamic graph sequences. Anomaly type, risk level, comprehensive score, key contribution factors, and root cause labels are used as supervision signals. The model aggregates spatial information through graph neural network layers. The evolution of node states within a continuous time window is learned through the GRU temporal coding layer: Finally, the diagnostic output layer outputs: in, This includes anomaly type, risk level, comprehensive safety and reliability score, key contribution factors, root cause candidates, and risk propagation paths. During online diagnosis, the system updates node features in real time according to a sliding time window and inputs them into the trained model to achieve safety and reliability status identification, short-term risk prediction, and interpretable diagnostic output.

9. The dynamic diagnostic system for the production safety and reliability of a chemical plant according to claim 1, characterized in that: The comprehensive safety and reliability evaluation and handling feedback module outputs operational status, anomaly type, risk level, comprehensive safety and reliability score, Top-K key contribution impact factors, risk propagation path, root cause candidate ranking, and handling suggestions. It pushes the diagnostic results to the operator interface for confirmation of receipt, handling, and review. Each anomaly diagnosis forms an event segment sample, which includes the anomaly start and end time, anomaly number, impact factor change, key contribution impact factor, root cause candidate, confirmed root cause, handling action, recovery time, false alarm cause, recurrence identifier, and review conclusion. The event segment samples are then fed back into the impact factor library, sample library, knowledge graph library, model parameter library, threshold library, and rule library. The comprehensive safety and reliability evaluation and handling feedback module follows these steps. Calculate the overall security and reliability score for the current time window, where Let i be the current overall weight of the i-th influence factor. Let i be the degree of abnormality of the i-th influencing factor. This is a penalty item for safety barrier failure. This is a penalty item based on the severity of the abnormal event. Bonus points for handling and restoring the status quo.

10. A method for dynamic diagnosis of the production safety and reliability of a chemical plant using the system described in any one of claims 1 to 9, characterized in that, Includes the following steps: S1. Parse multi-source files and real-time data to form structured source units; generate candidate impact factors based on the source paths and perform data entry gating; S2. Generate initial weights for impact factors based on source evidence; construct a knowledge graph of impact factors among equipment, parameters, impact factors, safety barriers, abnormal states, and evaluation results; generate a relation weight matrix based on the graph relationships; S3. Input the graph node features, relation weight matrix and time window sequence into the graph-enhanced spatiotemporal AI diagnostic module; S4. Generate dynamic contribution weights based on AI diagnostic results; integrate source evidence weights, graph relationship weights, dynamic contribution weights, and feedback correction weights to generate the current comprehensive weights; S5. Calculate the comprehensive safety and reliability score based on the current comprehensive weight and the degree of abnormality of the impact factors; S6. Feed the results of on-site confirmation, handling execution, restoration confirmation, false alarm correction and review back into the impact factor library, sample library, knowledge graph library, model parameters, threshold library and rule library.

Images