A chemical project safety design management method based on big data

Abstract

This invention discloses a big data-based safety design management method for chemical projects, relating to the field of big data technology in chemical safety. The method includes: collecting multi-source heterogeneous data and preprocessing it to form a project baseline dataset; collecting design change data and performing differential comparison with a safety knowledge graph to identify differential objects; starting from the differential objects, performing dependency propagation analysis along the safety dependencies in the safety knowledge graph to output affected objects and risk propagation chains; determining the review order of affected objects based on the risk propagation chains, sequentially extracting the specifications to be reviewed that match the affected objects from the specification knowledge base, determining whether the affected objects meet the specifications to be reviewed, and outputting the specification review conclusion. This invention realizes dynamic and precise management and automated compliance review of chemical project safety design, shifting from experience-driven to big data-driven management.

Description

Technical Field

[0001] This invention relates to the field of big data technology for chemical safety, and in particular to a big data-based method for safety design and management of chemical projects. Background Technology

[0002] Against the backdrop of digital transformation in the process industry, safety design management for chemical projects has shifted from traditional experience-driven to data-driven approaches. Existing chemical design management technologies primarily rely on the deep integration of computer-aided design systems and building information modeling (BIM) technology. By establishing static 3D physical models and attribute databases, geometric conflict detection and preliminary quantity surveying are achieved during the design process. These technologies typically utilize relational databases to store equipment parameters, process flow (P&ID) data, and material balance sheets, and attempt to perform consistency checks on design documents through pre-defined logical rules. With the development of big data management technologies, some systems have begun to introduce distributed storage architectures. These architectures aim to centrally manage massive amounts of historical design data, accident case databases, and standard specification documents, providing designers with retrieval assistance and playing a significant role in improving design efficiency and standardization.

[0003] However, existing technologies lack in-depth analysis of semantic relationships across disciplines and systems, making it difficult to achieve accurate risk linkage analysis when facing design changes. Specifically, chemical project design is extremely complex, and changes to a single parameter often trigger chain reactions through pressure balance, thermodynamic cycles, or control logic. Existing systems mostly rely on manual triggering of review processes or simple searches based on keyword matching, failing to automatically identify the hidden risk propagation paths behind changes. This often results in local optimizations causing global risks during design iterations, and the lack of intelligent decision support for review priorities during compliance reviews affects the rigor and timeliness of safety design. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a big data-based method for safety design management of chemical projects to solve the problems of lagging change risk linkage analysis and low efficiency of compliance review.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a big data-based method for safety design and management of chemical projects, which includes: collecting multi-source heterogeneous data and preprocessing the multi-source heterogeneous data to form a project baseline dataset; Entity and relation extraction are performed on the project baseline dataset according to the predefined chemical safety ontology framework to obtain entity objects and entity relationships. Based on the entity objects and entity relationships, a safety knowledge graph is generated through node mapping and edge mapping. Collect design change data and perform differential comparison between the design change data and the safety knowledge graph to identify the differential objects; Starting with the differential object, dependency propagation analysis is performed along the security dependency relationships in the security knowledge graph, and the affected objects and risk propagation chains are output. The review order of affected objects is determined according to the risk propagation chain. The standards to be reviewed that match the affected objects are extracted from the standard knowledge base in turn. It is then determined whether the affected objects meet the standards to be reviewed, and the standard review conclusion is output.

[0007] As a preferred embodiment of the big data-based chemical project safety design and management method of the present invention, the specific steps of collecting multi-source heterogeneous data and preprocessing the multi-source heterogeneous data to form a project baseline dataset are as follows: Perform missing value completion, outlier detection, and redundancy removal on multi-source heterogeneous data, and output cleaned data; The cleaned data is standardized in terms of units, terms, and versions. It is then associated and structured according to the topological hierarchy of the chemical project to output the project baseline dataset.

[0008] As a preferred embodiment of the big data-based chemical project safety design and management method of the present invention, the specific steps of extracting entities and relationships from the project baseline dataset according to a predefined chemical safety ontology framework to obtain entity objects and entity associations are as follows: Extract the entity category set and relation category set from the chemical safety ontology framework; Candidate entities are identified from the project baseline dataset based on the entity category set, and the candidate entities are disambiguated by name and encoded with unique identifiers to obtain entity objects; Based on the set of relationship categories, semantic analysis is used to identify hierarchical subordinate relationships, physical connection relationships, process constraint relationships, and security dependency relationships between entity objects in the project baseline dataset, thereby obtaining entity association relationships.

[0009] As a preferred embodiment of the big data-based chemical project safety design and management method described in this invention, the chemical safety ontology framework is constructed by performing entity recognition, relation learning, and ontology hierarchical constraint modeling on text in the chemical safety field using natural language processing methods.

[0010] As a preferred embodiment of the big data-based chemical project safety design and management method described in this invention, the generation of a safety knowledge graph through node mapping and edge mapping refers to mapping entity objects as graph nodes, mapping entity relationships as graph edges, and performing consistency verification and conflict resolution on graph nodes and graph edges to generate a safety knowledge graph.

[0011] As a preferred embodiment of the big data-based chemical project safety design management method of the present invention, the specific steps of collecting design change data and performing differential comparison between the design change data and a safety knowledge graph to identify the differential objects are as follows: Collect design change data, and standardize the format and align the terminology of the design change data to form standard design change data; Extract the change object, change object identifier, change attribute value, change relationship and change type from standard design change data; Based on the object identifier, search for the target object corresponding to the object being changed in the security knowledge graph; When the change type is attribute modification and the changed attribute value is inconsistent with the attribute value of the target object, the changed object will be identified as an attribute modification class differential object; When the change type is relation modification and the changed association relationship is inconsistent with the entity association relationship corresponding to the target object, the changed object will be identified as a relation modification type differential object.

[0012] As a preferred embodiment of the big data-based chemical project safety design and management method of the present invention, the step of starting with differential objects and performing dependency propagation analysis along the safety dependency relationships in the safety knowledge graph to output affected objects and risk propagation chains is as follows: Security dependencies are extracted from the security knowledge graph and mapped to first-order logic formulas. The maximum likelihood estimation method is used to assign confidence weights to the first-order logic formula, and a Markov logic network is constructed based on the first-order logic formula and the confidence weights. The entity relationships and attribute values ​​of entity objects are mapped to baseline observation evidence, and the baseline observation evidence is used as a constraint to perform probabilistic inference in the Markov logic network to obtain the baseline conditional probability distribution of entity objects. By instantiating predicates, the change type, change association, and change attribute value of the difference object are mapped to real-time observation evidence, and the real-time observation evidence is injected into the Markov logic network as a mandatory constraint node. Markov blankets are defined in the Markov logic network with mandatory constraint nodes as the center, and Gibbs sampling algorithm is used to calculate the real-time conditional probability distribution of entity objects within the Markov blankets. Calculate the relative entropy between the real-time conditional probability distribution and the baseline conditional probability distribution, and identify entities whose relative entropy exceeds the risk evolution threshold as affected entities; The Markov logic network is instantiated as a Markov random field, and the MaxWalkSAT algorithm is used to solve for the optimal truth set in the Markov random field. Starting with the differential object and ending with the affected object, the security dependencies that are true in the optimal truth set are sequentially connected into a directed dependency chain, and the directed dependency chain is defined as the risk propagation chain.

[0013] As a preferred embodiment of the big data-based chemical project safety design and management method described in this invention, the risk evolution threshold is set based on the statistical probability distribution of historical accident cases and the safety integrity level of the chemical project.

[0014] As a preferred embodiment of the big data-based chemical project safety design management method of the present invention, the specific steps of determining the review order of affected objects according to the risk propagation chain and sequentially extracting the specifications to be reviewed that match the affected objects from the specification knowledge base are as follows: The topological distance between the affected object and the differential object is calculated based on the risk propagation chain, and the weighted product of the reciprocal of the topological distance and the relative entropy is defined as the risk impact intensity. Based on the order of risk impact intensity from high to low, the unique object identifier of the affected object is used as an index to retrieve the standard to be reviewed that matches the affected object in the standard knowledge base.

[0015] As a preferred embodiment of the big data-based chemical project safety design management method of the present invention, the specific steps for determining whether the affected object meets the specifications to be reviewed and outputting the specification review conclusion are as follows: The threshold conditions and logical constraints in the specifications to be reviewed are transformed into specification judgment formulas through semantic parsing, and the geometric boundaries corresponding to the specification judgment formulas are constructed in the feature space through manifold learning algorithm. The attribute values ​​of the affected objects in the security knowledge graph and their corresponding entity associations are mapped to feature coordinate points in the feature space, and the topological inclusion relationship between the feature coordinate points and the geometric boundaries is calculated. Based on the topological inclusion relationship, when the feature coordinate point is located within the geometric boundary, the affected object satisfying the specification to be reviewed is taken as the specification review conclusion; Based on topological inclusion relationships, when the feature coordinate point is located outside the geometric boundary, the affected object is considered to have violated the specification to be reviewed, which is taken as the specification review conclusion.

[0016] The beneficial effects of this invention are as follows: by constructing a chemical safety knowledge graph to realize the structured association of design elements, by using Markov logic networks to transform the risk evolution caused by design changes into probabilistic inference, by quantitatively assessing the risk propagation path of affected objects, and by combining manifold learning algorithms to transform normative constraints into geometric boundaries of feature space, the invention realizes dynamic and precise management and automated compliance review of chemical project safety design from experience-driven to big data management-driven. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a big data-based safety design and management method for chemical projects.

[0019] Figure 2 A flowchart for obtaining entity relationships.

[0020] Figure 3 A flowchart for identifying differential objects.

[0021] Figure 4 A flowchart for generating a security knowledge graph. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a big data-based method for safety design and management of chemical projects, including the following steps: S1. Collect multi-source heterogeneous data and preprocess the multi-source heterogeneous data to form the project baseline dataset.

[0026] S1.1 Perform missing value completion, outlier detection, and redundancy removal on multi-source heterogeneous data, and output the cleaned data.

[0027] Furthermore, missing values ​​are imputed in multi-source heterogeneous data through mean or median interpolation, based on... The principle is to perform outlier detection on multi-source heterogeneous data, identify and process abnormal observation points that deviate from the norm, perform redundancy removal on multi-source heterogeneous data through primary key matching, eliminate duplicate information and invalid redundancy, and output cleaned data.

[0028] It should be noted that multi-source heterogeneous data includes structured data, semi-structured data, and unstructured data generated during the design, construction, operation, and safety management of chemical projects.

[0029] S1.2. After cleaning, the data is standardized in terms of units, terms and versions, and then associated and structured according to the topological hierarchy of the chemical project to output the project baseline dataset.

[0030] Furthermore, the cleaned data is standardized using range transformation or Z-score normalization. Terminology alignment is performed on the cleaned data according to the standard dictionary of chemical engineering ontology. Version alignment is performed on the cleaned data through timestamp verification and version status bit matching. Topological hierarchical association mapping is performed according to the logical sequence and spatial affiliation of chemical process flow. Discrete records are mounted to the corresponding process units. Multi-table merging and index reconstruction are performed to generate the project baseline dataset.

[0031] It should be noted that the standard dictionary of chemical engineering ontology is constructed by performing natural language processing on original documents in the field of chemical safety and using ontology modeling tools (such as Protégé) to extract entity categories, attribute relationships and hierarchical logic.

[0032] S2. Extract entities and relationships from the project baseline dataset according to the predefined chemical safety ontology framework, obtain entity objects and entity relationships, and generate a safety knowledge graph based on entity objects and entity relationships through node mapping and edge mapping.

[0033] S2.1 Extract the entity category set and relation category set from the chemical safety ontology framework.

[0034] Furthermore, by calling ontology parsing libraries such as OWL API or RDFLib, the XML / RDF serialized files of the chemical safety ontology framework are structurally traversed to extract entity class definitions, attribute axiom constraints, and SubClassOf inheritance paths. This enables automated parsing of categories, attributes, and levels. Ontology classes that can be used as object types are extracted from the ontology class definitions of the chemical safety ontology framework, forming a set of entity categories. Relationship types, relationship directions, value constraints, and applicable scope that can connect two types of entity objects are extracted from the relationship definitions of the chemical safety ontology framework, forming a set of relationship categories.

[0035] It should be noted that the chemical safety ontology framework is constructed by performing entity recognition, relation learning, and ontology hierarchical constraint modeling on chemical safety texts using natural language processing methods. Specifically, the TF-IDF algorithm is applied to extract feature terms from chemical safety texts and define them as ontology classes; the co-occurrence matrix is ​​used to analyze the correlation strength between terms and identify the collaborative influence relationships between entities; and the ontology definition language (OWL) is used to hierarchically model these categories and relationships according to the functional logic of safety management, ultimately constructing the chemical safety ontology framework.

[0036] S2.2 Identify candidate entities from the project baseline dataset based on the entity category set, and perform name disambiguation and unique identifier encoding on the candidate entities to obtain entity objects.

[0037] Furthermore, based on the name, attribute features, encoding rules, and contextual constraints in the entity category set, candidate entities are located in the table fields, text records, and structured entries of the project baseline dataset, and the name, number, location, time, and attribute values ​​are extracted from the data records corresponding to the candidate entities. For candidate entities with the same or similar names, the hierarchical level, associated objects, attribute combinations, and time stamps are compared to complete name disambiguation. A unique object identifier is generated for each unique entity according to the unified encoding rules, and the entity category, attribute value, and unique object identifier are recorded together as an entity object.

[0038] S2.3. Based on the set of relationship categories, semantic analysis is used to identify hierarchical subordinate relationships, physical connection relationships, process constraint relationships, and security dependency relationships between entity objects in the project baseline dataset, thereby obtaining entity association relationships.

[0039] Furthermore, based on the semantic definitions, directions of action, and applicable conditions of various relationships in the relationship category set, semantic analysis is performed on the text descriptions, field associations, topological hierarchy records, and process flow records in the project baseline dataset to extract relationship clues between entity objects. Semantic expressions indicating hierarchical affiliation are identified as hierarchical subordinate relationships; semantic expressions indicating spatial connectivity, pipeline connections, and interface correspondence are identified as physical connection relationships; semantic expressions indicating process parameter limitations, process sequence, and operating conditions are identified as process constraint relationships; and semantic expressions indicating risk triggering, protection constraints, and impact transmission are identified as safety dependency relationships. When the relationship category, relationship direction, and entity objects at both ends of the relationship are all confirmed to be consistent, the corresponding relationship is recorded as an entity association relationship.

[0040] It should be noted that semantic definition refers to the linguistic definition and logical description of the essential relationship between chemical entities, used to specify the specific connotation of the relationship. For example, "belonging to" is defined as a spatial or managerial hierarchical inclusion relationship between two objects.

[0041] The direction of action refers to the vector orientation of entity relationships in the graph structure. It is used to clarify the starting point and ending point of risk or logical flow. For example, in the security dependency relationship of "A triggers B", the direction of action specifies that it points from the risk source node to the affected node.

[0042] The applicable conditions refer to the constraints, attribute restrictions, or process prerequisites that must be met when establishing a specific relationship. For example, "process constraint relationship" only applies to process units that are in the same production process segment and have direct parameter coupling.

[0043] S2.4 Generating a security knowledge graph through node mapping and edge mapping refers to mapping entity objects as graph nodes, mapping entity relationships as graph edges, and performing consistency checks and conflict resolution on graph nodes and graph edges to generate a security knowledge graph.

[0044] Furthermore, based on the unique object identifier and attribute set of the entity object, corresponding graph nodes are created in the graph database, and the attribute values ​​are stored as node attributes; based on the relationship category, direction, and object identifiers at both ends in the entity association relationship, graph edges with specific semantic labels are constructed between the corresponding graph nodes; for graph nodes and graph edges, it is searched to see if there are historical nodes with the same unique object identifier but conflicting attributes, and the attribute conflict is resolved by attribute weight voting method, and isolated edges pointing to non-existent nodes are checked for integrity and removed; transitive closure operation is performed to detect and handle redundant paths with logical conflicts, and a security knowledge graph is generated.

[0045] S3. Collect design change data and perform differential comparison between the design change data and the safety knowledge graph to identify the differential objects.

[0046] S3.1 Collect design change data, and standardize the format and align the terminology of the design change data to form standard design change data.

[0047] Furthermore, the source format, field name, date format, numerical unit, and encoding form of each design change data entry are analyzed. Synonymous fields are mapped to unified field names, different units are converted to unified dimensions, and date and version information are organized into a unified record format. Based on the chemical industry ontology standard dictionary, the object names, attribute names, and relationship names in the design change data are aligned to eliminate inconsistencies in names and ambiguities in expression, forming standard design change data with consistent field structure and consistent terminology.

[0048] S3.2 Extract the change object, change object identifier, change attribute value, change association and change type from the standard design change data.

[0049] Furthermore, change entries are read at the record granularity in the standard design change data. The specific objects that have changed are extracted from the object name, object number, attribute, relationship, and change description fields of the standard design change data, and these are designated as the change objects. The change object identifier is extracted from the number, tag, or unique code fields of the standard design change data. The changed attribute value is extracted from the new value field of the attribute item in the standard design change data. The changed relationship is extracted from the description of newly added, deleted, or adjusted relationships in the standard design change data. The change type is determined based on the modification markers, field differences, and relationship differences in the change description field. For example, modification markers such as "added," "deleted," or "replaced" are identified in the change description field. Field differences are identified by comparing the numerical fluctuations of table fields before and after the change, and relationship differences are identified by verifying the addition or deletion of node connection relationships in the topology logic. Thus, the change entry is classified into specific change types such as attribute modification, entity addition, or relationship cancellation.

[0050] S3.3. Based on the changed object identifier, search for the target object corresponding to the changed object in the security knowledge graph.

[0051] Furthermore, a precise search is performed in the security knowledge graph based on the changed object identifier, and the changed object identifier is matched item by item with the unique object identifier of the entity object. When a completely identical unique object identifier exists, the name, attribute value and entity association relationship of the entity object are read from the node record and edge record of the corresponding target object in the security knowledge graph, and the target object is identified. When multiple approximate results appear, the object name, its level and associated context are compared again, and the unique matching result is retained, so that the changed object and the target object in the security knowledge graph are established to correspond.

[0052] S3.4 When the change type is attribute modification and the changed attribute value is inconsistent with the attribute value of the target object, the changed object will be identified as an attribute modification class differential object.

[0053] Furthermore, when the change type is determined to be attribute modification, the changed attribute value of the changed object is read from the attribute change record of the corresponding change item in the standard design change data, and the attribute value of the same attribute item is extracted from the attribute record of the target object in the safety knowledge graph. A consistency comparison is performed according to the attribute name, unit of measurement, value format and version status. When there is a substantial difference between the changed attribute value and the attribute value of the target object in terms of numerical value, text content or status expression, the changed object is marked as an attribute modification differential object.

[0054] S3.5 When the change type is relation modification and the changed association relationship is inconsistent with the entity association relationship corresponding to the target object, the changed object will be identified as a relation modification type differential object.

[0055] Furthermore, when the change type is determined to be a relationship modification, the change association of the changed object is read from the relationship change record of the corresponding change item in the standard design change data. The relationship category, relationship direction, and object identifiers involved at both ends of the relationship are parsed. The relationship of similar entities is extracted from the relationship record corresponding to the target object in the safety knowledge graph as a comparison benchmark. When the change association and the entity association corresponding to the target object are inconsistent in terms of relationship category, relationship direction, or associated object, the changed object is marked as a relationship modification differential object.

[0056] S4. Starting with the differential object, perform dependency propagation analysis along the security dependency relationships in the security knowledge graph, and output the affected objects and risk propagation chains.

[0057] S4.1 Extract security dependencies from the security knowledge graph and map the security dependencies into first-order logic formulas.

[0058] Furthermore, all entity relationships categorized as security dependencies are selected from the security knowledge graph, and a unified organization table is established based on the relationship start point, relationship end point, relationship category, relationship direction, triggering condition, and constraint attributes in the relationship records. A corresponding first-order logic formula description template is defined for each type of security dependency based on the unified organization table. The relationship start point and relationship end point are mapped to object items in the first-order logic formula, and the triggering condition, risk state, protection state, and constraint state are mapped to predicate items in the first-order logic formula. The relationship direction is written into the order of predicate parameters, and the applicable scope and hierarchical constraints are written into the applicable condition description of the first-order logic formula. Each organized security dependency is then converted into a structurally consistent first-order logic formula record, and the first-order logic formula is output.

[0059] It should be noted that the scope of application refers to the validity boundary of the risk evolution rule defined by the first-order logic formula in chemical plants, process links or specific operating conditions. For example, the scope of application of the "pressure chain reaction" rule is limited to high-pressure operating units such as hydrogenation reactors.

[0060] Hierarchical constraints refer to the chemical system topology or logical hierarchy that rules must follow to take effect during the propagation of safety dependencies. For example, it may stipulate that risk propagation can only occur within the same process branch, or that it must be passed up from the lower-level instrumentation sensing layer to the control logic layer.

[0061] S4.2. The maximum likelihood estimation method is used to assign confidence weights to the first-order logic formula, and a Markov logic network is constructed based on the first-order logic formula and the confidence weights.

[0062] Furthermore, in order to obtain the statistical samples required for calculating the confidence weights, historical project records, accident records, rectification records, and operation records that are semantically consistent with the first-order logic formula are collected. Terminology alignment, object identification unification, and status value standardization are performed on the historical project records, accident records, rectification records, and operation records so that the historical project records, accident records, rectification records, and operation records can correspond to the predicate terms in the first-order logic formula one by one. For each first-order logic formula, the frequency of observations where the result is true and the constraint is satisfied when the corresponding condition of the first-order logic formula occurs is counted, as well as the frequency of observations where the result is false or the constraint is not satisfied (for example, in the verification of safety rules for hydrogenation reactors, if the historical record shows that the pressure exceeds the design limit but the interlock vent valve is not opened, this record is counted as the frequency of observations where the result is false or the constraint is not satisfied). The proportion of the frequency of observations where the result is true and the constraint is satisfied in all corresponding observation frequencies is determined as the initial support of the first-order logic formula. Then, the initial support is reversed according to the frequency of observations where the result is false or the constraint is not satisfied to obtain the confidence weight corresponding to the first-order logic formula. The confidence weight is written into the confidence weight field of the corresponding first-order logic formula. A unified network file is constructed based on first-order logic formulas, confidence weights, object constant sets, predicate sets, and constraint sets. First-order logic formulas are used as network rules, confidence weights are used as network rule strengths, and object constants are substituted into network rules to generate computable rule instance structures, thus obtaining a Markov logic network.

[0063] The confidence weights of a first-order logic formula can be expressed as: ; in, Represents the confidence weights of a first-order logic formula; This indicates the frequency of observations in which the result is true and the constraints are satisfied; This indicates a smoothing term used to avoid zero values ​​in the numerator or denominator; This indicates the frequency of observations where the result is invalid or the constraint is not met.

[0064] S4.3 Map entity relationships and attribute values ​​of entity objects to baseline observation evidence, and use the baseline observation evidence as constraints to perform probabilistic inference in the Markov logic network to obtain the baseline conditional probability distribution of entity objects.

[0065] Furthermore, the entity relationships and attribute values ​​of entity objects in the security knowledge graph are read, and hierarchical subordinate relationships, physical connection relationships, process constraint relationships and security dependency relationships are mapped into relational benchmark observation evidence according to predefined predicate mapping rules. The status attributes, equipment attributes, process attributes and protection attributes of entity objects are mapped into attribute-type benchmark observation evidence, and each benchmark observation evidence is supplemented with object identifier, attribute name, attribute value, relationship category and version status. All baseline observation evidence is imported into the Markov logic network as fixed observation conditions. Under the condition that the fixed baseline observation evidence remains unchanged, the unobserved variables are set as variables to be inferred. For each variable to be inferred, the rule instances related to the variable to be inferred are read one by one. Based on the confidence weight of the relevant rule instances and the current values ​​of the other variables, the conditional probability corresponding to each state value of the variable to be inferred is calculated, and the current state probability of the variable to be inferred is updated with the calculation results. Repeat the above conditional probability calculation and state probability update for all variables to be inferred until the state probability changes of each variable to be inferred tend to stabilize. Then, normalize the conditional probabilities corresponding to each state value of each variable to be inferred to obtain the baseline conditional probability distribution corresponding to each entity object.

[0066] It should be noted that the predicate mapping rules are set based on the consistent correspondence between the entity relationships and attribute structures of entity objects in the security knowledge graph and the predicate semantics, parameter positions and constraints in the first-order logic formula.

[0067] S4.4. Through predicate instantiation, the change type, change association, and change attribute value of the difference object are mapped to real-time observation evidence, and the real-time observation evidence is injected into the Markov logic network as a mandatory constraint node.

[0068] Furthermore, the change types, change relationships, and change attribute values ​​of the difference objects are standardized and encoded. Based on the same predicate mapping rules as the baseline observation evidence, attribute modification-type difference objects are mapped to attribute state change predicates, relationship modification-type difference objects are mapped to relationship change predicates, change types are mapped to change state predicates, change attribute values ​​are mapped to attribute value predicates, and change relationships are mapped to relationship establishment or relationship cancellation predicates. For each piece of real-time observation evidence, object identifier, relationship category, attribute name, attribute value, change direction, and version status are supplemented. The real-time observation evidence is written into the variable position corresponding to the difference object in the Markov logic network, and the real-time observation evidence is marked as a mandatory constraint node that cannot be rewritten by the inference result, thus completing the injection of real-time observation evidence.

[0069] S4.5. Using the mandatory constraint nodes as the center, delineate the Markov blanket in the Markov logic network, and use the Gibbs sampling algorithm to calculate the real-time conditional probability distribution of entity objects within the Markov blanket.

[0070] Furthermore, taking the differential object variable corresponding to the mandatory constraint node as the center, all predicate instances that have direct logical association with the mandatory constraint node, adjacent variables that appear together with the predicate instances in the same first-order logic formula, and adjacent first-order logic formulas that have a constraint effect on adjacent variables are retrieved in the Markov logic network. Based on the retrieval results, the Markov blanket range of the mandatory constraint node is defined, and the variables, rules and observation conditions within the Markov blanket range are extracted into local inference subnetworks. In the local inference subnet, fixed mandatory constraint nodes and benchmark observation evidence are set as unobserved variables to be sampled. The unobserved variables are set as unobserved variables to be sampled. The unobserved variables are repeatedly conditionally sampled according to the Gibbs sampling algorithm. The state occurrence of the unobserved variables during the sampling process is recorded until the state frequency of each unobserved variable tends to stabilize. The real-time conditional probability distribution of the state variables related to each entity object is calculated based on the stable state frequency.

[0071] It should be noted that the Gibbs sampling algorithm is a random sampling method based on Markov chain Monte Carlo theory. It is used to generate sample sequences in high-dimensional probability distributions to approximate the target joint distribution or to calculate complex integral expectations. The principle is to use all conditional probability distributions of the target distribution for alternating sampling. In each iteration, the values ​​of other random variables are fixed and the values ​​of individual variables are updated according to the latest conditional probability distribution. As the number of iterations increases, the generated sample sequence will converge to the target joint distribution.

[0072] S4.6 Calculate the relative entropy between the real-time conditional probability distribution and the baseline conditional probability distribution, and identify entities whose relative entropy exceeds the risk evolution threshold as affected objects.

[0073] Furthermore, the real-time conditional probability distribution is aligned with the baseline conditional probability distribution based on the same entity object, the same state variable, and the same state value. Missing items are uniformly padded with zeros, and the order of state values ​​is consistent. When there is a state value in the baseline conditional probability distribution that is still zero after padding but whose corresponding real-time conditional probability distribution is not zero, the baseline probability corresponding to that state value is replaced with a preset minimum non-zero smoothed value. After completing the alignment and smoothing, the real-time probability and baseline probability corresponding to each state value under the same state variable are read for each entity object. The deviation of the real-time probability from the baseline probability is accumulated using the real-time probability corresponding to each state value as a weight, and the accumulated result is determined as the relative entropy of the corresponding entity object. The relative entropy of all entity objects is compared with the risk evolution threshold item by item, and entity objects whose relative entropy exceeds the risk evolution threshold are selected and the affected objects are output.

[0074] The relative entropy of an entity object can be expressed as: ; in, Represents the relative entropy of an entity object; This represents the index tag corresponding to a single state value; It represents the set of all possible values ​​for the same state variable; Indicates the state value index is Real-time probability at that time; Indicates the state value index is The baseline probability at that time, and This is the baseline probability after smoothing.

[0075] It should be noted that the minimum non-zero smoothing value is set based on the probability resolution of the state values ​​in the baseline conditional probability distribution, the statistical precision of the samples, and the need to perform stabilization processing on zero-probability values.

[0076] The risk evolution threshold is set based on the statistical probability distribution of historical accident cases and the safety integrity level of the chemical project. Its example value range is... The example range is based on the following: analysis of the statistical probability distribution of historical chemical accident cases and the safety integrity level of chemical projects, combined with the sensitivity of affected objects in the risk propagation chain and the severity of accident consequences, to determine the acceptable risk deviation limit.

[0077] S4.7 Instantiate the Markov logic network into a Markov random field, and use the MaxWalkSAT algorithm in the Markov random field to solve for the optimal truth set.

[0078] Furthermore, the object constants, predicate instances, and rule instances in the Markov logic network are converted into variable nodes and potential function structures in the Markov random field. The rule satisfaction situation corresponding to the first-order logic formula is converted into the truth consistency requirement of the variable node combination. The confidence weight is converted into the preference intensity of the corresponding potential function, and the values ​​of the observed variables in the baseline observation evidence and the real-time observation evidence are kept constant. In Markov random fields, unobserved variables are set as search variables. The solution objectives are to maximize rule satisfaction, minimize conflicting rules, and maximize total weighted consistency. Currently unsatisfied rule instances are randomly selected. From the search variables involved in these unsatisfied rule instances, the improvement in rule satisfaction and the change in total weighted consistency after truth flipping are tested one by one. The search variable that maximizes the improvement in rule satisfaction and the highest increase in total weighted consistency after truth flipping is selected for truth flipping. When multiple search variables have the same improvement result, one search variable is randomly selected from the multiple search variables for truth flipping. Repeat the process of selecting unsatisfied rule instances, testing the variables to be searched, and flipping truth values ​​until all rule instances are satisfied, or until the total weighted consistency no longer improves after multiple iterations. Then, determine the truth value configuration of the variable with the highest total weighted consistency as the optimal truth value group.

[0079] It should be noted that the MaxWalkSAT algorithm is a heuristic search algorithm for solving weighted Boolean satisfaction problems. It is used to find near-optimal solutions under large-scale logical constraints. The principle is to randomly select an unsatisfied constraint and alternate between local search and random jumps to increase the number of satisfied constraints and avoid getting trapped in local optima.

[0080] S4.8 Starting with the differential object and ending with the affected object, sequentially connect the security dependencies that are true in the optimal truth set into a directed dependency chain, and define the directed dependency chain as a risk propagation chain.

[0081] Furthermore, starting with the differential object and ending with the affected object, all security dependency instances with true values ​​are extracted from the optimal truth set, and a traceable connection table is established according to the relationship direction, object identifier, and hierarchical constraints. Path backtracking is performed on each affected object in the traceable connection table to retrieve the continuous security dependency sequence that starts from the differential object and eventually reaches the affected object. Security dependency instances that cannot be closed to the affected object are removed, and security dependency instances that can form continuous propagation are connected sequentially according to the occurrence order and dependency direction to form a directed dependency chain from the differential object to the affected object. All directed dependency chains are then organized and defined as risk propagation chains.

[0082] S5. Determine the review order of affected objects based on the risk propagation chain, extract the standards to be reviewed that match the affected objects from the standard knowledge base in turn, determine whether the affected objects meet the standards to be reviewed, and output the standard review conclusion.

[0083] S5.1 Calculate the topological distance between the affected object and the differential object based on the risk propagation chain, and define the risk impact intensity as the weighted product of the reciprocal of the topological distance and the relative entropy.

[0084] Furthermore, all directed dependency chains from the differential object to each affected object in the risk propagation chain are read. For each directed dependency chain, the connection level of the continuous security dependency relationship between the differential object and the affected object is counted. The connection level corresponding to the directed dependency chain with the fewest connection levels is determined as the topological distance between the affected object and the differential object. Simultaneously, the relative entropy obtained by the affected object in the dependency propagation analysis stage is read. The relative entropy is standardized in terms of dimensions and normalized in intervals to eliminate the difference in value scale between different affected objects. According to the preset weight configuration table, the reciprocal of the topological distance and the normalized relative entropy are weighted and combined to obtain the risk impact intensity corresponding to each affected object.

[0085] It should be noted that the weighting table is set based on the topological distance of the affected objects, the statistical impact of their relative entropy in historical projects and safety incidents, and the safety integrity requirements of chemical projects. For example, higher risk impact weights are assigned to affected objects with shorter topological distances and higher relative entropy.

[0086] S5.2. Following the order of risk impact intensity from high to low, use the unique object identifier of the affected object as an index to retrieve the standard to be reviewed that matches the affected object from the standard knowledge base.

[0087] Furthermore, all affected objects are sorted in descending order according to the risk impact intensity in the risk impact intensity record table to form a review sequence of affected objects. Using the unique object identifier of the affected object as the primary search key, a joint search is performed in the corresponding object index table, object category index table, applicable scenario index table, and version status index table in the specification knowledge base to extract the specification entries corresponding to the object category, process location, equipment purpose, attribute name, entity association category, and version status that are consistent with the affected objects. The specification entries are organized into a list of specifications to be reviewed that correspond one-to-one with the affected objects, and the specifications to be reviewed are output.

[0088] It should be noted that the standard knowledge base is a structured database formed by knowledge extraction and semantic modeling of original documents in the chemical industry. It includes an object index table, an object category index table, an applicable scenario index table, and a version status index table.

[0089] S5.3. The threshold conditions and logical constraints in the specifications to be reviewed are transformed into specification judgment formulas through semantic parsing, and the geometric boundaries corresponding to the specification judgment formulas are constructed in the feature space through manifold learning algorithm.

[0090] Furthermore, the standard text in the standard to be reviewed is segmented into sentences, aligned with terms, and matched with semantic templates. Statements containing attribute names, comparison relationships, and restricted values ​​are extracted as threshold conditions, and statements containing object categories, entity relationship categories, relationship directions, and constraint requirements are extracted as logical constraints. The threshold conditions and logical constraints are then organized into a standard judgment formula according to a unified field structure (including object category, applicable scenario, attribute name, comparison relationship, restricted value, entity relationship category, relationship direction, and constraint status). Based on the standard judgment formula, historical object samples with consistent object categories, applicable scenarios, and version status are selected from the security knowledge graph and historical project records. Historical object samples that meet the standards to be reviewed and those that violate the standards to be reviewed are extracted separately. Attribute features and entity relationship features corresponding to the standard judgment formula are extracted from the historical object samples. The attribute features are subjected to unified dimensionality, state standardization, and unified coding. The entity relationship features are subjected to relationship category alignment, relationship direction unification, and hierarchical position coding. The processed attribute features and entity relationship features are combined into a sample feature vector. A manifold learning algorithm is used to map the sample feature vectors to a two-dimensional feature space, so that historical object samples that meet the review criteria and those that violate the review criteria form distinguishable distribution regions in the two-dimensional feature space. For the two-dimensional feature coordinate points corresponding to the historical object samples that meet the review criteria, the Alpha Shape algorithm is used to extract the geometric boundaries. Specifically, Delaunay triangulation is constructed for the two-dimensional feature coordinate points, triangles with circumcircle radii greater than the preset rolling sphere radius are deleted, and the boundary points on the outer boundary of the remaining triangles are retained as the outermost contour points. The outermost contour points are then connected end to end according to the boundary connection order to form a closed polygon. This closed polygon is determined as the geometric boundary corresponding to the criterion judgment formula.

[0091] It should be noted that the rolling ball radius is set based on the nearest neighbor distance statistics of the two-dimensional feature coordinate points corresponding to the historical object samples that meet the specifications to be reviewed.

[0092] Manifold learning algorithms are nonlinear dimensionality reduction methods that recover low-dimensional manifold structures from high-dimensional sampled data. They are used for dimensionality reduction, feature extraction, and data visualization of high-dimensional data. The principle is to assume that high-dimensional observation data are distributed on low-dimensional manifolds embedded in high-dimensional space and to attempt to map high-dimensional data to low-dimensional space by preserving local geometric structure or global topological features.

[0093] S5.4 Map the attribute values ​​of the affected objects in the security knowledge graph and their corresponding entity associations to feature coordinate points in the feature space, and calculate the topological inclusion relationship between the feature coordinate points and the geometric boundaries.

[0094] Furthermore, attribute values ​​and entity relationships corresponding to the affected objects are extracted from the security knowledge graph. The attribute values ​​and entity relationships involved in the judgment are selected according to the standard judgment formula corresponding to the standard to be reviewed. The attribute values ​​involved in the judgment are subjected to dimensional unification, state standardization and unified coding. The entity relationships involved in the judgment are subjected to relationship category alignment, relationship direction unification and hierarchical position coding. The processed attribute values ​​and entity relationships are combined into object feature vectors according to the same judgment item order, the same feature item order and the same coding rules used when constructing the geometric boundary. The object feature vectors are mapped to feature coordinate points in the two-dimensional feature plane through the same manifold learning mapping rules as when constructing the geometric boundary. The geometric boundary is considered as a closed polygon formed by sequentially connecting the boundary points end to end. Based on the feature coordinates in the two-dimensional feature plane, a decision ray is drawn along a fixed direction. The intersection of the decision ray with each side of the closed polygon is calculated one by one. When the number of valid intersections between the decision ray and each side of the closed polygon is odd, the feature coordinates are determined to be inside the geometric boundary. When the number of valid intersections between the decision ray and each side of the closed polygon is even, the feature coordinates are determined to be outside the geometric boundary. When the feature coordinates coincide with any boundary point or are located on any boundary edge, the feature coordinates are determined to be on the geometric boundary, and being on the geometric boundary is considered as being inside the geometric boundary. Based on the determination results of whether the feature coordinate point is located inside the geometric boundary, outside the geometric boundary, or on the geometric boundary, the topological inclusion relationship between the feature coordinate point and the geometric boundary is obtained.

[0095] S5.5 According to the topological inclusion relationship, when the feature coordinate point is located within the geometric boundary, the affected object that meets the specification to be reviewed shall be taken as the specification review conclusion.

[0096] S5.6 According to the topological inclusion relationship, when the feature coordinate point is located outside the geometric boundary, the affected object is considered to have violated the specification to be reviewed as the specification review conclusion.

[0097] In summary, this invention achieves structured association of design elements by constructing a chemical safety knowledge graph, transforming the risk evolution caused by design changes into probabilistic inference using Markov logic networks, quantitatively assessing the risk propagation path of affected objects, and combining manifold learning algorithms to transform normative constraints into geometric boundaries of the feature space. This enables dynamic and precise management and automated compliance review of chemical project safety design, moving from experience-driven to big data-driven.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A big data-based chemical project safety design management method, characterized in that, include: Collect multi-source heterogeneous data and preprocess the multi-source heterogeneous data to form the project baseline dataset; Entity and relation extraction are performed on the project baseline dataset according to the predefined chemical safety ontology framework to obtain entity objects and entity relationships. Based on the entity objects and entity relationships, a safety knowledge graph is generated through node mapping and edge mapping. Collect design change data and perform differential comparison between the design change data and the safety knowledge graph to identify the differential objects; Starting with the differential object, dependency propagation analysis is performed along the security dependency relationships in the security knowledge graph, and the affected objects and risk propagation chains are output. The review order of affected objects is determined according to the risk propagation chain. The standards to be reviewed that match the affected objects are extracted from the standard knowledge base in turn. It is then determined whether the affected objects meet the standards to be reviewed, and the standard review conclusion is output.

2. The big data-based chemical project safety design management method according to claim 1, characterized in that, The process of collecting multi-source heterogeneous data and preprocessing it to form a project baseline dataset involves the following steps: Perform missing value completion, outlier detection, and redundancy removal on multi-source heterogeneous data, and output cleaned data; The cleaned data is standardized in terms of units, terms, and versions. It is then associated and structured according to the topological hierarchy of the chemical project to output the project baseline dataset.

3. The big data-based chemical project safety design management method according to claim 2, characterized in that, The steps for extracting entities and relationships from the project baseline dataset according to the predefined chemical safety ontology framework are as follows: Extract the entity category set and relation category set from the chemical safety ontology framework; Candidate entities are identified from the project baseline dataset based on the entity category set, and the candidate entities are disambiguated by name and encoded with unique identifiers to obtain entity objects; Based on the set of relationship categories, semantic analysis is used to identify hierarchical subordinate relationships, physical connection relationships, process constraint relationships, and security dependency relationships between entity objects in the project baseline dataset, thereby obtaining entity association relationships.

4. The big data-based chemical project safety design management method according to claim 3, characterized in that, The chemical safety ontology framework is constructed by performing entity recognition, relation learning, and ontology hierarchical constraint modeling on text in the chemical safety field using natural language processing methods.

5. The method for safety design and management of chemical projects based on big data as described in claim 1, characterized in that, The process of generating a security knowledge graph through node mapping and edge mapping refers to mapping entity objects as graph nodes, mapping entity relationships as graph edges, and performing consistency checks and conflict resolution on graph nodes and graph edges to generate a security knowledge graph.

6. The method for safety design and management of chemical projects based on big data as described in claim 5, characterized in that, The specific steps for collecting design change data and comparing it with a safety knowledge graph to identify discrepancies are as follows: Collect design change data, and standardize the format and align the terminology of the design change data to form standard design change data; Extract the change object, change object identifier, change attribute value, change relationship and change type from standard design change data; Based on the object identifier, search for the target object corresponding to the object being changed in the security knowledge graph; When the change type is attribute modification and the changed attribute value is inconsistent with the attribute value of the target object, the changed object will be identified as an attribute modification class differential object; When the change type is relation modification and the changed association relationship is inconsistent with the entity association relationship corresponding to the target object, the changed object will be identified as a relation modification type differential object.

7. The method for safety design and management of chemical projects based on big data as described in claim 6, characterized in that, Starting with the differential object, dependency propagation analysis is performed along the security dependency relationships in the security knowledge graph to output the affected object and the risk propagation chain. The specific steps are as follows: Security dependencies are extracted from the security knowledge graph and mapped to first-order logic formulas. The maximum likelihood estimation method is used to assign confidence weights to the first-order logic formula, and a Markov logic network is constructed based on the first-order logic formula and the confidence weights. The entity relationships and attribute values ​​of entity objects are mapped to baseline observation evidence, and the baseline observation evidence is used as a constraint to perform probabilistic inference in the Markov logic network to obtain the baseline conditional probability distribution of entity objects. By instantiating predicates, the change type, change association, and change attribute value of the difference object are mapped to real-time observation evidence, and the real-time observation evidence is injected into the Markov logic network as a mandatory constraint node. Markov blankets are defined in the Markov logic network with mandatory constraint nodes as the center, and Gibbs sampling algorithm is used to calculate the real-time conditional probability distribution of entity objects within the Markov blankets. Calculate the relative entropy between the real-time conditional probability distribution and the baseline conditional probability distribution, and identify entities whose relative entropy exceeds the risk evolution threshold as affected entities; The Markov logic network is instantiated as a Markov random field, and the MaxWalkSAT algorithm is used to solve for the optimal truth set in the Markov random field. Starting with the differential object and ending with the affected object, the security dependencies that are true in the optimal truth set are sequentially connected into a directed dependency chain, and the directed dependency chain is defined as the risk propagation chain.

8. The method for safety design and management of chemical projects based on big data as described in claim 7, characterized in that, The risk evolution threshold is set based on the statistical probability distribution of historical accident cases and the safety integrity level of chemical projects.

9. The method for safety design and management of chemical projects based on big data as described in claim 7, characterized in that, The step of determining the review order of affected objects based on the risk propagation chain, and sequentially extracting the relevant standards to be reviewed from the standard knowledge base, is as follows: The topological distance between the affected object and the differential object is calculated based on the risk propagation chain, and the weighted product of the reciprocal of the topological distance and the relative entropy is defined as the risk impact intensity. Based on the order of risk impact intensity from high to low, the unique object identifier of the affected object is used as an index to retrieve the standard to be reviewed that matches the affected object in the standard knowledge base.

10. The method for safety design and management of chemical projects based on big data as described in claim 9, characterized in that, The specific steps for determining whether the affected object meets the specifications to be reviewed and outputting the specification review conclusion are as follows: The threshold conditions and logical constraints in the specifications to be reviewed are transformed into specification judgment formulas through semantic parsing, and the geometric boundaries corresponding to the specification judgment formulas are constructed in the feature space through manifold learning algorithm. The attribute values ​​of the affected objects in the security knowledge graph and their corresponding entity associations are mapped to feature coordinate points in the feature space, and the topological inclusion relationship between the feature coordinate points and the geometric boundaries is calculated. Based on the topological inclusion relationship, when the feature coordinate point is located within the geometric boundary, the affected object satisfying the specification to be reviewed is taken as the specification review conclusion; Based on topological inclusion relationships, when the feature coordinate point is located outside the geometric boundary, the affected object is considered to have violated the specification to be reviewed, which is taken as the specification review conclusion.

Images