Risk analysis agent system and method based on cooperation of large model and early warning model

Abstract

The invention provides a risk analysis agent system and method based on cooperation of a large model and an early warning model, and the method comprises the steps: analyzing a natural language request into a device identifier, a time range and an index field, and generating a specification containing a query specification, a risk hypothesis set and a verification criterion through a large model which is finely adjusted through a field instruction; task arrangement forms a preferred model path, candidate mapping and a directional replanning rule; acquiring data according to the query specification, calculating slope and mean value features, judging according to a verification criterion, and generating determined or candidate hypothesis identifiers; an interpretation and disposal report is generated according to the hypothesis identification retrieval knowledge base and aligned with the query specification, and a major hazard source early warning interpretation result is output; the method ensures that the device, the time and the field boundary are consistent, data boundary drift caused by path switching is avoided, judgment is repeatable, a result is traceable, and the consistency of early warning interpretation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of safety risk analysis of dangerous chemical production, and particularly relates to a risk analysis method and system based on cooperation of a large model and an early warning model. BACKGROUND

[0002] Dangerous chemical production device operation involves multi-source monitoring data, historical trends and process safety knowledge, and field early warning relies on accurate reference and interpretation of device identification, time range and index. Business personnel often make analysis requests in natural language, expecting to be converted into executable structured queries and enter the early warning model evaluation, while combining failure mechanisms and evidence patterns to generate disposable reports. Due to sampling intervals, units and field naming, and frequent scene changes, maintaining the consistency of data boundaries and terms becomes a basic requirement for risk analysis automation. In the management of major hazard sources, it is required that the interpretation results be traceable, reviewable, and able to be associated with existing knowledge bases and process specifications in a closed loop. Therefore, the end-to-end connection from natural language to models and knowledge, as well as the automatic arrangement of stable boundaries, has become a common concern in the industry.

[0003] Existing technologies mostly use rule dictionaries or template-driven natural language parsing to map devices, times and indexes in the request into structured queries. After obtaining historical and real-time data, quantitative evaluation is performed according to preset early warning models or algorithm libraries, and risk scores are generated. Some solutions combine knowledge bases and process specifications to match mechanism explanations and disposal suggestions for abnormal features; there are also process tools that use arrangement engines to realize data pulling, model calling and report generation, supporting field unit conversion, sampling alignment and index-level display, forming regular early warning and explanation output for major hazard sources. In terms of implementation, common practices include merging aliases through word tables, standardizing time expressions, and constructing SQL or time series query instructions; on the model side, threshold determination, trend fitting and rule base comparison are mainly used, and report templates are used to output explanation content. The arrangement system usually supports the calling and switching of multiple model paths, and completes data acquisition, evaluation and documentation in steps, and finally delivers with device and time labels.

[0004] However, the existing solutions generally lack a closed-loop organization centered on candidate failure mechanisms-evidence patterns-early warning model paths, and there is a lack of stable anchors between evaluation and explanation. Secondly, model path switching is mainly based on strategy or operator adaptation, which is easy to cause implicit changes in device, time and index boundaries, leading to data boundary drift and traceability difficulties. Thirdly, the judgment rules and feature extraction methods are not unified and fixed, making it difficult to ensure consistent reference and repeatable judgment of slope, mean value and other features in different links.

[0005] Therefore, a production risk analysis method and system that can solve the above-mentioned deficiencies of the prior art is needed to solve the problems of those skilled in the art. SUMMARY

[0006] An object of the present application is to provide a risk analysis method and system based on large model and early warning model cooperation. In the risk analysis scene of dangerous chemical production, a cooperative closed loop from natural language request to data query, early warning model evaluation and knowledge explanation is constructed, so that the evaluation and explanation always converge around the candidate failure mechanism, and the device identifier, time range and index field boundary are strictly locked in the whole process of path switching, avoiding data boundary drift, ensuring that the judgment and report can be directly traced to the query specification and stably delivered in the major hazard source early warning explanation; the specific requirements unify the reference of feature calculation and verification criteria, solidify the binding relationship of query specification and assumption identifier, so that natural language, model path and knowledge base entries are arranged and executed within a single boundary, thereby realizing end-to-end consistency and traceability in actual device scenarios.

[0007] According to the risk analysis method based on large model and early warning model cooperation, the method comprises the following steps:

[0008] S1, receiving a natural language task request, analyzing the dangerous chemical production device identifier, time range and monitoring index, and forming an analysis target description;

[0009] S2, generating a risk analysis specification using a large language model based on the analysis target description, the risk analysis specification including query specification, risk hypothesis set and verification criteria, the query specification limiting the device identifier, time range and index field, the risk hypothesis set giving the candidate failure mechanism, corresponding evidence mode and its early warning model path label, and the verification criteria being used for judging the consistency of the quantitative evaluation result and the evidence mode, and outputting the risk analysis specification;

[0010] S3, generating a task execution plan object by a risk analysis task arrangement intelligent agent based on the risk analysis specification, the task execution plan object including a preferred model path, a candidate hypothesis mapping table and a directional re-planning rule, the directional re-planning rule specifying that when the verification criteria are not met, the model path is switched according to the candidate hypothesis mapping table and the query specification remains unchanged, so that the evaluation and explanation converge around the risk hypothesis, and outputting the task execution plan object;

[0011] S4, generating a structured query instruction and obtaining a target data object according to the query specification and the task execution plan object, and keeping the field and time boundary consistent under the constraint of the directional re-planning rule;

[0012] S5, calculating a quantitative evaluation result based on the target data object according to the preferred model path, and judging based on the verification criteria, generating a determined assumption identifier when the criteria are met, and generating a candidate assumption identifier when the criteria are not met;

[0013] S6, retrieving the knowledge base based on the hypothesis identification and the quantitative evaluation result, and generating an explanation and treatment report object with the hypothesis identification as an anchor point;

[0014] S7, outputting the explanation and treatment report object as a major hazard source early warning explanation result.

[0015] Optionally, step S1 is specifically:

[0016] The natural language task request is taken as input, text normalization and sentence segmentation are performed, an analysis channel oriented to dangerous chemical production device identification, time range and monitoring index is established, and modifiers and redundant descriptions irrelevant to the task are removed;

[0017] The dangerous chemical production device identification is identified from the natural language task request, the reference and alias are merged, and a single and referable dangerous chemical production device identification is determined;

[0018] The time range is parsed, the start time and the end time are identified, the absolute time and the relative time are unified, and the time range containing the start time and the end time is formed;

[0019] The monitoring index is parsed and mapped to the index field, the unit of measurement and the sampling interval are unified, and the analysis target description is assembled based on the determined dangerous chemical production device identification, time range and monitoring index;

[0020] The analysis target description is format checked to ensure that the dangerous chemical production device identification is unique, the time range contains the start time and the end time, the monitoring index corresponds to the index field and the units are consistent, and the analysis target description can be directly referenced.

[0021] Optionally, step S2 is specifically:

[0022] The analysis target description is taken as input, the domain instruction fine-tuning large language model is called, the three-part structure of the risk analysis specification is generated, the query specification, the risk hypothesis set and the verification criterion are explicitly included, and the reference relationship between the query specification, the risk hypothesis set and the verification criterion is established;

[0023] The query specification is generated according to the device identification, the time range and the monitoring index in the analysis target description, the device identification, the time range and the index field are limited, and item-by-item mapping and unification are performed, so that the query specification can be directly used for structured query instruction;

[0024] The risk hypothesis set is generated, the candidate failure mechanism is taken as a unit, the corresponding evidence mode is determined, and the early warning model path label is bound for each candidate failure mechanism, so that the candidate failure mechanism, the evidence mode and the early warning model path label form a parameterized and referable set;

[0025] The verification criterion is generated according to the evidence mode in the risk assumption set, the judgment boundary and the control relationship are set, the consistency determination mode of the quantitative evaluation result and the evidence mode is specified, and the consistent and inconsistent determination output is given;

[0026] The consistency of the risk analysis specification is checked, the consistency of the terms and the closed reference of the query specification, the risk assumption set and the verification criterion are checked, and the constraint relationship between the warning model path label and the device identifier, the time range and the index field is confirmed;

[0027] The pre-constraint of the warning model path label and the verification criterion is introduced in the risk analysis specification generation stage, the above constraint is taken as the input starting point of the task execution plan object, and the task arrangement and interpretation are ensured to converge around the candidate failure mechanism;

[0028] The query specification, the risk assumption set and the verification criterion are solidified in a unified format, the boundary consistency of the device identifier, the time range and the index field is maintained, and the risk analysis specification is output.

[0029] Optionally, step S3 is specifically:

[0030] The risk analysis specification is taken as input, the query specification, the risk assumption set and the verification criterion are parsed by the risk analysis task arrangement intelligent agent, the device identifier, the time range and the index field in the query specification are extracted, the candidate failure mechanism, the evidence mode and the warning model path label in the risk assumption set are extracted, the judgment boundary and the control relationship in the verification criterion are extracted, and the field mapping relationship between the above content and the task execution plan object is established;

[0031] Based on the warning model path label in the query specification and the risk assumption set, the preferred model path is determined, the warning model path label with the completely consistent index field set in the query specification is preferentially selected, if there is no completely consistent one, the one with the largest number of required index fields is selected, and the binding mode of the device identifier, the time range and the index field is solidified in the preferred model path;

[0032] A candidate assumption mapping table is constructed, the verification criterion is decomposed into judgment items according to the index field, each judgment item is specified with a candidate failure mechanism and its warning model path label, the evidence mode corresponding to the specified warning model path label is required to directly involve the index field corresponding to the judgment item, and the switching direction of the candidate failure mechanism and the preferred model path is recorded;

[0033] Directional re-planning rules are generated, the judgment result of the verification criterion is taken as the trigger condition, when any judgment item does not meet the condition, the corresponding warning model path label is switched according to the candidate assumption mapping table, the query specification is kept unchanged, the device identifier, the time range and the index field are prohibited to be modified, and the judgment boundary and the control relationship of the verification criterion are kept unchanged;

[0034] The switching sequence control is set in the directional re-planning rule, the switching is triggered one by one according to the listed order of the judgment items in the verification criterion, parallel triggering is prohibited, and the output definition of the quantitative evaluation result in the task execution plan object is used after the switching;

[0035] The hypothesis identification generation rule is preset in the task execution plan object, the candidate failure mechanism is one-to-one corresponding to the early warning model path label to form the hypothesis identification, and the pointing of the hypothesis identification to be generated is updated when the path is switched, so that the hypothesis identification can be consistent with the query specification;

[0036] The preferred model path, the backup hypothesis mapping table and the directional re-planning rule are assembled into the task execution plan object, and the task execution plan object is output.

[0037] Optionally, step S4 is specifically:

[0038] The query specification and the task execution plan object are taken as inputs, the device identification, the time range and the index field in the query specification are parsed, the preferred model path and the directional re-planning rule in the task execution plan object are parsed, and the binding mode in the task execution plan object is used as a constraint for generating the structured query instruction;

[0039] The structured query instruction is generated according to the query specification, the device identification is limited to the device identification in the query specification, the time boundary is limited to the time range in the query specification, and the field is limited to the index field in the query specification, thereby forming the structured query instruction which can be directly executed;

[0040] The structured query instruction is executed to obtain the target data object, when the directional re-planning rule triggers the path switching, the query specification is kept unchanged according to the directional re-planning rule, the structured query instruction is repeatedly executed, and it is ensured that the device identification, the time range and the index field are consistent with the query specification;

[0041] The consistency of the target data object is checked, the index field of the target data object is checked to cover the index field in the query specification, the time boundary of the target data object is checked to cover the time range in the query specification, and it is confirmed that the target data object can be directly referenced by the preferred model path or the early warning model path after the switching.

[0042] Optionally, step S5 is specifically:

[0043] The target data object and the preferred model path are taken as inputs, the target data object is quantitatively calculated according to the preferred model path, and the quantitative evaluation result is generated;

[0044] The consistency of the quantitative evaluation result is determined according to the judgment boundary and the contrast relationship in the verification criterion, and the determination result of meeting or not meeting is obtained;

[0045] When the consistency determination is satisfied, the candidate failure mechanism is combined with the early warning model path label to generate a determined hypothesis identification, and the determined hypothesis identification is consistently referenced with the device identification, the time range, and the index field;

[0046] When the consistency determination is not satisfied, the candidate failure mechanism is combined with the early warning model path label to generate a candidate hypothesis identification, and the candidate hypothesis identification is consistently referenced with the device identification, the time range, and the index field.

[0047] Optionally, step S6 is specifically:

[0048] The hypothesis identification and the quantitative evaluation result are taken as inputs, the knowledge base is searched with the hypothesis identification as an anchor point, the search range is limited to the candidate failure mechanism and the evidence mode corresponding to the hypothesis identification, and the boundaries of the device identification, the time range, and the index field in the query specification are used;

[0049] According to the evidence mode corresponding to the hypothesis identification in the knowledge base, the judgment boundary and the contrast relationship in the verification criterion are used to contrast the quantitative evaluation result, and a contrast conclusion of mechanism consistency or inconsistency is formed;

[0050] According to the contrast conclusion, the corresponding relationship of the hypothesis identification, the quantitative evaluation result, and the evidence mode is written into the explanation and disposal report object, and the correspondence with the device identification, the time range, and the index field in the query specification is explicitly stated in the explanation and disposal report object;

[0051] According to the corresponding relationship of the knowledge base and the hypothesis identification, the contrast conclusion is merged to generate the explanation and disposal report object, which is marked with the device identification, the time range, and the index field in the query specification, and the explanation and disposal report object is output.

[0052] Optionally, step S7 is specifically:

[0053] The explanation and disposal report object is aligned with the query specification, the device identification, the time range, and the index field in the query specification are used for marking, the hypothesis identification and the quantitative evaluation result are taken as reference items of the explanation and disposal report object, and the corresponding explanation and disposal content is formed based on the judgment conclusion of the verification criterion, and the candidate failure mechanism and the early warning model path label corresponding to the hypothesis identification are listed in the explanation and disposal report object;

[0054] The explanation and disposal report object is subjected to consistency checking, the hypothesis identification and the quantitative evaluation result are confirmed to be consistent with the device identification, the time range, and the index field in the query specification, it is confirmed that the device identification, the time range, and the index field do not change in the two cases of triggered and not triggered directional re-planning rules, and the consistency checking conclusion is retained;

[0055] In the explanation and disposal report object, the determined hypothesis identifier or the candidate hypothesis identifier is marked, and the determination conclusion of the early warning model path label and the verification criterion are associated, so that each item of content can be directly traced by the query specification, while maintaining consistent reference of the device identifier, the time range and the index field, and the explanation and disposal report object is output as the major hazard source early warning explanation result.

[0056] A risk analysis intelligent agent system based on large model and early warning model cooperation, comprising:

[0057] A natural language analysis module for analyzing dangerous chemical production device identifier, time range and monitoring index, and generating analysis target description;

[0058] A risk analysis specification generation module for calling field instruction fine-tuning large language model to generate query specification, risk hypothesis set and verification criterion;

[0059] A risk analysis task scheduling module for generating a task execution plan object containing a preferred model path, a candidate hypothesis mapping table and a directional re-planning rule;

[0060] A query and data acquisition module for generating a structured query instruction and acquiring a target data object according to the query specification and the task execution plan object;

[0061] A quantitative evaluation and hypothesis identification module for calculating quantitative evaluation results according to the preferred model path and generating a determined hypothesis identifier or a candidate hypothesis identifier according to the verification criterion;

[0062] A knowledge retrieval and explanation generation module for retrieving a knowledge base with the hypothesis identifier as an anchor point and generating an explanation and disposal report object according to the evidence mode and the quantitative evaluation results;

[0063] An alignment and output module for aligning the explanation and disposal report object with the query specification and consistency checking, associating the early warning model path label with the determination conclusion, and outputting the major hazard source early warning explanation result.

[0064] The beneficial effects of the present application are:

[0065] 1. The proposal presents an improved risk analysis task scheduling and targeted re-planning method based on the ternary organization of "candidate failure mechanism-evidence mode-early warning model path", using the technical means of preferred model path selection, candidate hypothesis mapping table and unmodifiable field set, and triggering single-step switching control with the decision items of verification criteria; the preferred path is determined by the adaptation score of coverage priority and extension penalty, the binding mode of device identifier, time range and index field is fixed during the switching process, and the defined output is used; the improved method keeps the data boundary stable during the whole process of model path switching, makes the evaluation and interpretation always converge around the failure mechanism, avoids data boundary drift, and ensures consistent reference and traceability of report and query specifications, meeting the stable delivery requirements of on-site major hazard source early warning interpretation.

[0066] 2. The proposal presents a new three-stage risk analysis report generation method and hypothesis identification mechanism, which uses large language models with field instruction fine-tuning to output closed structures containing query specifications, risk hypothesis sets and verification criteria, explicitly binds early warning model path labels and decision trigger relationships during the generation phase, and completes the verification of term consistency, reference closure and boundary locking constraints; technical means include unified executable description of slope and mean characteristics, sampling interval alignment and weighted consistency score, and hypothesis identification composed of "candidate failure mechanism + path label" (carrying device, time and field meta-information); the method ensures the repeatability of mechanism determination and the accuracy of field-level comparison by keeping the decision rules consistent in analysis, evaluation and interpretation, and provides stable anchors for knowledge retrieval and disposal generation.

[0067] 3. The proposal presents a collaborative method from natural language parsing to structured query, quantitative evaluation and knowledge retrieval alignment, which generates structured query instructions by locking the boundary of query specifications, obtains target data objects by combining time coverage, field coverage and sampling consistency verification, and drives knowledge base retrieval and disposal report generation with hypothesis identification as anchor, finally outputs major hazard source early warning interpretation results aligned with query specifications. The overall method ensures that usable data enters the evaluation through the data consistency score threshold, performs hypothesis identification meta-information verification, quantitative result boundary verification and path switching stability verification in the report stage, forming an end-to-end referenceable closed loop. Unlike the existing method of flow splicing based on template or strategy adaptation, the proposal takes boundary stability and mechanism anchoring as the core design, so that natural language, model path and knowledge base entries are scheduled and executed within a single boundary, solving the problems of collaborative closed loop and traceability consistency from the scene perspective. BRIEF DESCRIPTION OF DRAWINGS

[0068] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation of the application. In the drawings:

[0069] Figure 1 A flow chart of a risk analysis method based on large model and early warning model cooperation is proposed for the present application;

[0070] Figure 2 A natural language analysis flow chart of a risk analysis method based on large model and early warning model cooperation is proposed for the present application;

[0071] Figure 3 A risk analysis specification generation flow chart of a risk analysis method based on large model and early warning model cooperation is proposed for the present application;

[0072] Figure 4 A task arrangement and directional re-planning flow chart of a risk analysis method based on large model and early warning model cooperation is proposed for the present application;

[0073] Figure 5 A structured query and data acquisition flow chart of a risk analysis method based on large model and early warning model cooperation is proposed for the present application;

[0074] Figure 6 A quantitative evaluation and judgment flow chart of a risk analysis method based on large model and early warning model cooperation is proposed for the present application;

[0075] Figure 7 A directional re-planning before and after comparison view of a risk analysis agent system based on large model and early warning model cooperation is proposed for the present application;

[0076] Figure 8 A system integration overview screenshot of a production risk analysis system based on large model and early warning model cooperation is proposed for the present application. DETAILED DESCRIPTION

[0077] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0078] REFERENCE Figures 1 to 8 A risk analysis method based on large model and early warning model cooperation, characterized in that it comprises the following steps:

[0079] S1, receiving a natural language task request, analyzing the dangerous chemical production device identification, time range and monitoring indicators, and forming an analysis target description;

[0080] S2, according to the analysis target description, using the field instruction to fine-tune the large language model to generate the risk analysis specification, the risk analysis specification includes query specification, risk hypothesis set and verification criterion, the query specification limits device identification, time range and index field, the risk hypothesis set gives candidate failure mechanism, corresponding evidence mode and its early warning model path label, the verification criterion is used for judging the consistency of quantitative evaluation result and evidence mode, and the risk analysis specification is output;

[0081] S3, based on the risk analysis specification, a task execution plan object is generated by a risk analysis task scheduling intelligent agent, the task execution plan object includes a preferred model path, a candidate hypothesis mapping table and a directional re-planning rule, the directional re-planning rule specifies that when the verification criterion is not met, the model path is switched according to the candidate hypothesis mapping table and the query specification remains unchanged, so that the evaluation and explanation converge around the risk hypothesis, and the task execution plan object is output;

[0082] S4, according to the query specification and the task execution plan object, a structured query instruction is generated and a target data object is obtained, and the field and time boundary are kept consistent under the constraint of the directional re-planning rule;

[0083] S5, based on the target data object, the quantitative evaluation result is calculated according to the preferred model path, and based on the verification criterion, the determined hypothesis identifier is generated when the verification criterion is met, and the candidate hypothesis identifier is generated when the verification criterion is not met;

[0084] S6, based on the hypothesis identifier and the quantitative evaluation result, the knowledge base is retrieved with the hypothesis identifier as the anchor point, and the explanation and disposal report object are generated.

[0085] In this embodiment, step S1 is specifically:

[0086] The natural language task request is recorded as Before entering the analysis, the text normalization and sentence segmentation are performed on Character standardization is adopted to convert full-width characters into half-width characters, Chinese numbers are unified into Arabic numbers, unit expressions are unified into standard unit words, and full stop, semicolon and comma are used as segmentation boundaries to form a sentence segment sequence;

[0087] An analysis channel is established for dangerous chemical production device identification, time range and monitoring index, and term matching and structure extraction are respectively performed in the corresponding channel, and irrelevant modifiers and redundant descriptions are removed at the channel entrance, so that subsequent extraction only acts on the fragments directly related to dangerous chemical production device identification, time range and monitoring index;

[0088] In the resolution of hazardous chemical production equipment identification, the sentence sequence is input into the hazardous chemical production equipment identification resolution channel, and the hazardous chemical production equipment identification vocabulary is used for matching. Candidate segments are extracted and the main segment is determined according to the longest matching priority strategy.

[0089] Construct an identification mapping table for hazardous chemical production facilities, merge designations and aliases, and map all candidate fragments to a unique identifier. ;

[0090] When multiple candidate segments are mapped to different identifiers, the identifier corresponding to the segment appearing in the first sentence is given priority, and subsequent segments are eliminated while maintaining the original identifier. only;

[0091] When multiple fragments are mapped to the same identifier, that identifier is retained. Remove duplicate segments to ensure that a single, referable identifier for a hazardous chemical production facility is identified.

[0092] In time range analysis, the absolute and relative times in the sentence segment sequence are unified;

[0093] The absolute time phrase is directly parsed into start time and end time, denoted as follows: and The relative time phrase is recorded using the system timestamp. Perform the conversion if the phrase is in the past. Minutes, then the time range is denoted as ;

[0094] When multiple time expressions exist, a candidate time range set is generated. Calculate the intersection ,like If not empty, then As a time range;

[0095] like If empty, then take the extended coverage area. As a time range;

[0096] This ultimately forms a time range that includes both start and end times. and the identification of hazardous chemical production facilities. Bind;

[0097] In the analysis of monitoring indicators, the nouns of monitoring indicators are extracted from the sentence sequence to form a set of monitoring indicators, denoted as . ;

[0098] Each monitoring indicator Mapped to system-referenceable indicator fields, forming a set of indicator fields denoted as . ;

[0099] The unit of each monitoring index is unified, the unit specification is normalized to the standard unit of the corresponding index field, and the sampling interval is unified to the sampling interval explicitly mentioned in the natural language task request. If it is not explicitly mentioned, the default sampling interval of the current monitoring index is taken as the unified value, denoted as ;

[0100] After the unit and sampling interval are unified, the dangerous chemical production device identifier , the time range and the index field set are assembled to form an analysis target description, ensuring the correspondence and referenceability between the contents;

[0101] In the format verification, the integrity and consistency of the analysis target description are verified;

[0102] The dangerous chemical production device identifier is verified for uniqueness, confirming that there is only one and consistent with the segment in the sentence segment sequence;

[0103] The time range is verified for boundary integrity, confirming that the time range contains the start time and the end time , and ;

[0104] The monitoring index is verified for field consistency, confirming that the monitoring index set and the index field set correspond, confirming that the unit of the index field after unit unification is the standard unit, and confirming that all index fields share the sampling interval after sampling interval unification ;

[0105] After the above verification is satisfied, the analysis target description can be directly referenced by the subsequent steps, and the dangerous chemical production device identifier, time range and index field in the analysis target description are kept consistent in the subsequent objects.

[0106] In this embodiment, step S2 is specifically:

[0107] The analysis target description is denoted as , which contains the device identifier , the time range and the index field set ;

[0108] Take as input, call the domain instruction fine-tuning large language model, and use the system prompt , domain instruction Structural constraints The combination requires the output to strictly follow a three-part structure, corresponding to the query specification, risk hypothesis set, and verification criteria respectively. During the generation stage, the reference links from the query specification to the risk hypothesis set and from the risk hypothesis set to the verification criteria should be explicitly established, so that the three can maintain consistent boundaries in terms of device identification, time range, and indicator fields and can reference each other.

[0109] When generating query specifications, with , and Perform item-by-item mapping for the input;

[0110] The device identifier is limited to Limit the time range to The start and end times in the data limit the indicator field to: All fields in;

[0111] Field names should follow the same naming rules as the indicator field set; units should follow the same unit rules as the indicator field set; and sampling intervals should use the sampling intervals already standardized in the analysis objective description. Write boundary locking constraints within the query specification to stipulate that the device identifier, time range, and indicator fields cannot be modified in subsequent objects and can only be called as references, so as to ensure consistency between the generation of subsequent structured query instructions and the task execution plan object;

[0112] When generating the risk hypothesis set, the set is constructed on a per-candidate failure mechanism basis;

[0113] Let each candidate failure mechanism be denoted as... In the set of indicator fields Screening and The subset of fields directly related to the mechanism is denoted as ;

[0114] Based on existing knowledge bases and process safety specifications The mechanism is developed to generate an evidence pattern, which clearly describes the expected characteristics, cross-comparison relationships and allowable deviations of each field over a time range.

[0115] for The path label of the bound early warning model is denoted as Agreement To identify the path that can be directly referenced by subsequent task execution plan objects, a parameterized representation is used. With the query specifications , and Perform boundary binding to form a set of referable and determinate risk hypotheses;

[0116] When generating verification criteria, the evidence model is transformed into decisionable rules oriented towards quantitative evaluation results, within a time frame. The slope and mean of the fields obtained from subsequent quantitative evaluation are used as features. A consistency score is defined for each candidate failure mechanism, and the following determination formula is adopted:

[0117] ;

[0118] in, For candidate failure mechanisms Consistency score, for The set of indicator fields involved To verify the fields in the criteria The weight, In the time range The slope of the field obtained by subsequent quantitative evaluation using least-squares linear fitting is shown above. For evidence pattern to fields The expected sign of the slope, For fields The slope threshold, In the time range The mean of the field is calculated by subsequent quantitative evaluation. This serves as a reference value for the mean of the evidence pattern. The mean deviation threshold, This is a decision function; the value is 0 if the condition is true.

[0119] Using consistency score threshold As a boundary for judgment, it is stipulated that when The output is consistent, when The output is inconsistent.

[0120] threshold , , Compared with reference value The weights are given by the knowledge base entries and process safety specifications. The sensitivity to the mechanism was set and normalized.

[0121] When performing consistency verification on the risk analysis statement, the query specifications, risk assumption set and verification criteria are compared item by item.

[0122] check Is it A subset of the data is used to verify whether each field has a corresponding decision item in the expected features of the evidence pattern, whether each decision item provides a clear threshold and reference value, whether the weights are normalized, and to verify each... Binding unique early-warning model path tags ;

[0123] After the check passes, record the reference closure relationship, ensuring that subsequent reading by the task execution plan object can directly point to specific fields and judgment items in the query specification and verification criteria;

[0124] Introduce the pre-constraint of early-warning model path tags and verification criteria in the risk analysis specification generation phase;

[0125] Each As the preferred model path or candidate model path in the task execution plan object, the consistency score The judgment output is used as the trigger condition for directional re-planning, and after triggering, the device identifier, time range, and index field in the query specification remain unchanged;

[0126] Through this pre-constraint, task arrangement and interpretation converge around the candidate failure mechanism, and data boundary drift is avoided during path switching;

[0127] Solidify the query specification, risk assumption set, and verification criteria in a unified format;

[0128] Solidify the combination of structured text paragraphs and key-value descriptions, ensuring that the device identifier , time range , and index field set appear consistently in the three-part structure and do not undergo alias replacement;

[0129] Ensure and the binding relationship between the judgment parameters is stable;

[0130] After solidification, output the risk analysis specification, which is specified as the unique input source for subsequent structured query instruction generation, task execution plan object construction, and consistency judgment;

[0131] Further, it is specified that the calculation method of slope and mean in the verification criteria is given in executable description;

[0132] The slope is calculated using least squares linear fitting of equally spaced samples within the time range , and the mean is calculated using the arithmetic mean within the time range :

[0133] When the sampling interval is not consistent, linear interpolation is performed with the sampling interval in the query specification as the target interval before calculation;

[0134] The above calculation method is taken as a component of the verification criterion, so that the quantitative evaluation result has a repeatable feature extraction method when entering the judgment;

[0135] The reference relationship is supplemented: the present application corresponds each index field in the query specification to a field subset in the risk hypothesis set, corresponds each field subset to a judgment item in the verification black, corresponds each candidate failure mechanism to a unique early warning model path label, and binds each early warning model path label to a trigger condition of the consistency score;

[0136] Through this organization, when the risk analysis specification is read by the task execution plan object, the preferred model path, the candidate hypothesis mapping table, and the input required by the directional re-planning rule can be directly determined, without the need to parse the natural language content again, ensuring that the query specification, the risk hypothesis set, and the verification criterion are consistent in the field and time boundary and can be directly used in the subsequent steps.

[0137] In the embodiment, step S3 is specifically:

[0138] The risk analysis specification is recorded as The risk analysis task scheduling intelligent agent parses , extracts the device identifier , the time range , and the index field set in the query specification, extracts the candidate failure mechanism set , the corresponding evidence mode, and the early warning model path label set in the risk hypothesis set, extracts the judgment boundary and the comparison relationship in the verification criterion, and establishes the field mapping relationship between the above content and the task execution plan object;

[0139] The task execution plan object is recorded as , the key-value binding , and is bound in , the index binding and , and the judgment item set in the verification criterion and its boundary parameters are bound in sequence;

[0140] When determining the preferred model path, the index field set in the query specification is compared with the field subset in the risk hypothesis set;

[0141] If there is an early warning model path label , the corresponding early warning model path label is set as the preferred model path recorded as , and The binding method between the intermediate curing device identification, time range, and indicator fields;

[0142] If there are no completely identical ones, then for all Calculate the fit score and select the early warning model path label with the highest fit score as the optimal choice. The following formula for the fitting score is used:

[0143] ;

[0144] in, Path labels for early warning models The fit score, To query the set of indicator fields in the specification, For candidate failure mechanisms The corresponding evidence pattern involves a subset of indicator fields. To verify the fields in the validation criteria In mechanism Weight on, This is the weighting coefficient for the coverage item, and it takes a positive value. This is the extensional penalty coefficient, which takes a non-negative value. The cardinality of the field difference set represents the path label of the early warning model. The number of fields required but not provided in the query specification;

[0145] With the maximum corresponding set up and solidify , and The binding method.

[0146] When constructing the candidate hypothesis mapping table, the verification criteria are decomposed into a set of decision items according to the indicator fields, denoted as _____. ;

[0147] Based on the judgment item For the index, a candidate failure mechanism and its early warning model path label are assigned to each decision item. The evidence corresponding to the specified early warning model path label must directly relate to the indicator field corresponding to that decision item, forming a candidate hypothesis mapping table denoted as […]. ;

[0148] exist Record from the preferred model path The switching direction of the path label to the backup early warning model is recorded, along with the trigger judgment item number of the indicator fields and verification criteria involved in the switching, to ensure that subsequent switching can be executed accurately.

[0149] When generating directional replanning rules, the judgment result of the verification criteria is used as the trigger condition;

[0150] When any decision item If not satisfied, refer to the candidate hypothesis mapping table. Switch to the corresponding early warning model path label, keep the query specifications unchanged, and do not modify the device identifier. Time range With indicator field set And keep the judgment boundaries and control relationships in the verification criteria unchanged;

[0151] Let the directional replanning rule be denoted as ,exist The document specifies the trigger source, switching target, set of unmodifiable fields, and decision boundary constraints.

[0152] Set switching sequence control in the directional replanning rules;

[0153] The switching should be triggered sequentially according to the order of the judgment items in the verification criteria, and parallel triggering is prohibited. After each switching is completed, the task execution plan object should be reused. The output definition of the quantitative evaluation results allows subsequent judgments to use them directly without introducing new output formats;

[0154] If multiple conditions are not met in a series of consecutive triggers, a sequential execution strategy of listing first and switching first is adopted. After each switch, the remaining decision items are re-evaluated to avoid boundary conflicts caused by multiple targets switching at the same time.

[0155] Pre-defined assumption identifier generation rules in the task execution plan object;

[0156] Candidate failure mechanisms Early warning model path labels Correspondingly, the generated hypothesis identifier is denoted as ;

[0157] When a directional replanning rule triggers a switch, the pointer of the hypothesis identifier to be generated is updated, so that... The path label is consistent with the currently used early warning model and matches the device identifier in the query specification. Time range With indicator field set Maintain consistent referencing;

[0158] Preferred model path Alternative Hypothesis Mapping Table With directional replanning rules Assembled into a task execution plan object ;

[0159] exist The system defines fixed field mapping relationships, fixed switching order control and unmodifiable field sets, fixed assumption identifier generation rules and path switching direction; China retains , , With weight The reference enables the quantitative evaluation results to be directly determined and driven by the verification criteria. The execution;

[0160] After assembly is complete, output the task execution plan object. It provides direct access for generating structured query commands and performing quantitative evaluation calculations.

[0161] In this embodiment, step S4 specifically includes:

[0162] Let the query specification be denoted as The task execution plan object is recorded as ;

[0163] Analysis Obtain device identification Time range Indicator Field Set With sampling interval ;

[0164] Analysis Obtain the preferred model path With directional replanning rules Continue to use The binding method in the middle will , and As an immutable constraint for generating structured query instructions;

[0165] The field names and units in the query specifications should follow the unified expressions already used in the risk analysis report to avoid adding aliases or changing units;

[0166] When generating structured query instructions, with The input construction instruction is denoted as ,exist The settings for the device filter are as follows: Set the time filter as start time With end time Set field projection to All fields in;

[0167] A uniform interval is used for sampling. For field values ​​with inconsistent sampling, linear interpolation is used to align the timestamp to the specified range without changing the value domain. ;

[0168] Prohibited in Add to Irrelevant filter options, ensuring that structured query commands are only affected by... , and constraint;

[0169] Execute structured query instructions The target data object is obtained and denoted as ;

[0170] Will Organization by field Indexed time series sets and uniform timestamp sequences Each of them Corresponding value sequence ;

[0171] When directional replanning rules Do not modify when path switching is triggered. Repeat execution and maintain , and Unchanged; only the path label of the currently used early warning model is updated in the task execution plan object for subsequent quantitative evaluation and interpretation, ensuring the stability of data boundaries and field sets;

[0172] For target data object To encapsulate, , , and Write as metadata, record exist The actual start and end times within are denoted as follows: and For each field Count the number of valid samples Statistical sample size for a unified timestamp sequence Calculate the interval between adjacent samples ;

[0173] The above statistics are only used for consistency verification and preparation for subsequent quantitative evaluation, and do not change the The value;

[0174] Consistency verification is jointly determined by time coverage, field coverage, and sampling consistency, and is calculated using a composite consistency score, defined as follows:

[0175] ;

[0176] in, The consistency score of the target data object. The actual start time of the target data object within the time range. The actual end time of the target data object within the time range. , , This is a weighting coefficient, taken as a non-negative value, and set in the task execution plan object. The cardinality of the indicator field set. For fields The number of valid samples within the time frame. To ensure a uniform number of samples in the timestamp sequence, To set the field coverage threshold, bind it to the indicator field in the risk analysis specification. The interval between adjacent timestamps To query the sampling interval of the specification, This is a decision function; it takes the value 1 if the condition is true and 0 if the condition is false.

[0177] Set a consistency threshold ,when The time determination is consistent, confirm. Preferred model path Alternatively, the path of the switched early warning model can be directly referenced;

[0178] when Inconsistent judgments are made when the same query specification is applied to the command. Perform a single re-execution and maintain , and constant;

[0179] Through the above generation, execution, and verification process, the structured query instructions are strictly constrained by the query specifications and task execution plan objects. This ensures that the boundaries of the device identifier, time range, and indicator fields remain consistent whether the directional replanning rule is triggered or not. This provides directly referable target data objects for the subsequent calculation, interpretation, and generation of disposal report objects for quantitative evaluation results.

[0180] In this embodiment, step S5 specifically includes:

[0181] Let the target data object be denoted as Let the preferred model path be denoted as ;

[0182] Identified by the query specification reading device Time range Indicator Field Set With sampling interval , keep the field naming and units consistent with the risk analysis specification;

[0183] will be organized as a time series collection indexed by the indicator fields, where each field corresponds to a value sequence and a uniform timestamp sequence, both within the time range and aligned to the sampling interval ;

[0184] will be loaded as a computation pipeline, with its field mapping and feature operators, prohibited from modifying , and boundaries;

[0185] will be

[0186] calculated for each field using the least squares linear fitting within the time range to obtain the slope feature , and for each field using the arithmetic mean calculation within the time range to obtain the mean feature , keeping the sample interception and sequence length consistent with the sampling interval ;

[0187] will be and summarized to form the quantitative evaluation result, denoted as , where is composed of a set of key-value pairs, with the key being the field and the value being a feature pair composed of and for direct reference in subsequent consistency determination;

[0188] will be

[0189] determined for consistency according to the judgment boundaries and control relationships in the verification criteria; , with the one-to-one corresponding candidate failure mechanism recorded as , to compare the threshold value and reference value specified by the verification criteria, and determine based on the consistency score and threshold ;

[0190] will be recorded as satisfied when the consistency score is not less than the threshold, and as not satisfied when the consistency score is less than the threshold; ​​

[0191] For cases involving multiple decision items, the verification criteria are listed in order, and each item is compared and a single decision conclusion is synthesized. The decision items are not synthesized in parallel.

[0192] When the consistency determination is satisfied, the hypothesis identifier generation rule in the task execution plan object is used to identify the candidate failure mechanism. Early warning model path labels Combined generation of determined hypothesis identifiers, denoted as ;

[0193] exist Embedded device logo Time range With indicator field set The reference metadata ensures that the identified hypothesis identifiers are consistent with the boundaries of the query specifications and with the quantitative evaluation results. Maintain field-level traceability;

[0194] Will Used for anchoring the subject of subsequent interpretation and handling of the report, and is prohibited from [using] [the following]. , and Make any replacements or extensions;

[0195] When the consistency criteria are not met, the candidate hypothesis mapping table in the task execution plan object is read. Using the first non-metd requirement as an index, the corresponding candidate failure mechanism and its warning model path label are selected, denoted as follows: and ;

[0196] When multiple candidate failure mechanisms can be indexed by the same decision item, the candidate failure mechanism with the largest corresponding weight in the verification criterion is selected as the sole candidate.

[0197] Will and Combined to generate candidate hypothesis identifiers, denoted as and in Embedded device logo Time range With indicator field set The reference metadata ensures that the candidate hypothesis identifier is consistent with the query specification and can be directly referenced by the subsequent interpretation and processing report objects;

[0198] This application explains the linear relationship of the technical route. During the generation of the determined hypothesis markers and candidate hypothesis markers, the application does not quantify the evaluation results. Perform secondary processing without creating new feature sets;

[0199] All comparisons are strictly performed according to the judgment boundary and control relationship given by the verification criteria;

[0200] If the conclusion does not meet the inconsistent field of the candidate path of the candidate hypothesis mapping table, the candidate path is directly eliminated according to the constraint of the unmodifiable field set in the task execution plan object, and the next available candidate path is selected, until the candidate hypothesis identification is generated or the candidate path is exhausted;

[0201] Through the above calculation, determination and identification generation process, the determined hypothesis identification and the candidate hypothesis identification are consistent with the query specification in device identification, time range and index field boundary, and provide directly citable anchor points for subsequent interpretation and disposal report object retrieval and organization.

[0202] In the embodiment, step S6 is specifically:

[0203] Let the hypothesis identification be , and let the quantitative evaluation result be , let the query specification be , wherein contains device identification , time range and index field set ;

[0204] Let the knowledge base be , and in , establish an index with the candidate failure mechanism set and the corresponding evidence mode set , and keep one-to-one binding with the early warning model path label set ;

[0205] Take as input to analyze the corresponding mechanism and the early warning model path label , limit the search range to only entries, and use , in , and as boundaries, without expanding to other device identifications, time ranges and index fields;

[0206] In the retrieval stage, take as an anchor point to query , return the field subset of the evidence mode , the expected slope symbol set , the slope threshold set , the mean reference value set and the mean deviation threshold set ;

[0207] Perform boundary validation on the returned content to confirm. for A subset of the threshold set and the reference value set are both given in the verification criteria and can be directly used for judgment;

[0208] If found Including those not in If a field is specified in the search results, the corresponding entries for that field will be removed from the search results to ensure that subsequent comparisons only involve fields specified in the search results. Conducted within the boundaries;

[0209] In the control phase, based on the judgment boundaries and control relationships in the verification criteria, the... Perform a consistency check;

[0210] Read Each field Features ,in In the time range Slope characteristics calculated internally. In the time range The mean characteristics calculated within the range;

[0211] For each Two judgments are made: Judgment 1 is and The direction is compared with the threshold to determine the second one. and and Deviation comparison;

[0212] The consistency score of the candidate failure mechanism is obtained by synthesizing all decision items using the consistency score rule in the verification criteria. , and consistency score threshold When a comparison is performed, Not less than The timing is consistent with the mechanism, when Less than The timing is inconsistent with the mechanism;

[0213] During the report generation phase, the field-level correspondence between the comparison conclusions and the evidence patterns is written into the interpretation and disposition report object, denoted as... ;

[0214] exist China and Israel As an explanatory anchor, the following are listed ( )and For each corresponding item, the satisfaction status of the slope and mean criteria is marked for each field, and referenced in each entry. In , and To maintain the correspondence between device identification, time range, and indicator fields;

[0215] The knowledge base targeting The proposed solutions are attached as structured entries. Only select with For directly relevant processing items, it is forbidden to attach content that is unrelated to the current boundary, ensuring that the report object can be directly invoked by subsequent outputs;

[0216] In the conclusion merging stage, the comparative conclusions are summarized based on the correspondence between the knowledge base and hypothesis labels. When the mechanisms are consistent, the conclusions are... The mechanistic conclusions are marked as consistent, and the early warning model path labels are associated. This is for future tracing purposes, and in cases where the mechanisms are inconsistent, [the following will be considered]. The mechanistic conclusions are labeled as inconsistent, and at the same time... The specific fields and corresponding boundary parameters for which the condition is not met are retained, and extension to them is prohibited. Other fields;

[0217] Ultimately In , and Identify and output the objects of the explanation and handling report. ,ensure The device identification, time range, and indicator field boundaries are completely consistent with the query specifications and can be directly referenced by subsequent interpretation results of major hazard source warnings.

[0218] In this embodiment, step S7 specifically includes:

[0219] The recipients of the explanation and handling report are recorded as follows: The query specification is recorded as ,in Includes device identification Time range With indicator field set ;

[0220] Let the hypothesis be labeled as The quantitative evaluation results will be recorded as Let the set of path labels for the early warning model be denoted as ;

[0221] When performing alignment operations, The header of the page is set to , and Do not add or replace field aliases;

[0222] Will With as reference item written , ensure that the reference path directly points to the determined hypothesis identification or the complement hypothesis identification and its corresponding field characteristics;

[0223] According to the determination conclusion of the verification criterion, generate the explanation and disposal content in , list the candidate failure mechanism corresponding to the hypothesis identification and its early warning model path label, and use field-level mapping to juxtapose the field characteristics in with the corresponding evidence mode item, so that each mapping falls within the boundary of and and can be directly indexed;

[0224] When performing consistency check on the explanation and disposal report object, three verifications are performed in turn;

[0225] First, verify the meta-information carried by the hypothesis identification, confirm that the device identification, time range and index field set in are consistent with , and , if not consistent, mark the inconsistent source in and block the output;

[0226] Then verify the boundary consistency of the quantitative evaluation result, confirm that the field set involved in is consistent with , confirm that all data timestamps referred to in are within , and do not exceed the time range or miss the required fields;

[0227] Finally, verify the stability under the triggering and non-triggering of the directional re-planning rule, read the reference to the path switching state in , confirm that the path switching only changes the early warning model path label, and does not change , and . Record the consistency check conclusion as , express it in with a Boolean flag and a text description, where the Boolean flag reflects whether it passes or not, and the text description gives the field name or time segment that does not pass, ensuring that subsequent symmetric true direct reference can obtain the complete alignment state.

[0228] When performing final labeling in the explanation and disposal report object, distinguish based on the type of ;

[0229] If is a determined hypothesis identification, mark it as Then in The labels in the text have been identified and associated with the early warning model path tags. The judgment conclusion of the verification criteria is consistent, and the satisfaction status and corresponding boundary parameters of each indicator field are marked in the field-level comparison.

[0230] like The alternative hypothesis is denoted as Then in The alternative hypothesis identifier is marked and associated with the early warning model path label. The conclusion of the verification criteria cited is inconsistent, and the specific judgment items that are not met and their threshold sources are retained in the field-level comparison.

[0231] Both types of labels require that device markings be maintained. Time range With indicator field set exist The one-to-one correspondence in the text is prohibited from being extended to other forms. Other fields or splits The boundary;

[0232] After completing the annotation, Final labeling is performed according to query specifications, to , and As the sole external retrieval key, it ensures that the content of explanations and actions, hypothesis identification, and quantitative evaluation results can all be directly traced through the query specifications;

[0233] by As a basis for consistency, it allows subsequent objects to be directly determined without repeated verification. Availability;

[0234] Ultimately The output is the interpretation result of the early warning for major hazard sources, ensuring that the output object aligns with the boundaries of the device identification, time range, and indicator fields. Strict consistency, and can be directly referenced by subsequent processes for binding trigger conditions for display, delivery and disposal execution.

[0235] 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 risk analysis method based on large model and early warning model cooperation, characterized in that, Comprising the following steps: S1, receiving a natural language task request, parsing the dangerous chemical production device identifier, time range and monitoring index, forming an analysis target description; S2, according to the analysis target description, using the field instruction fine-tuning large language model to generate the risk analysis specification, the risk analysis specification contains query specification, risk hypothesis set and verification criterion, the query specification limits the device identifier, time range and index field, the risk hypothesis set gives the candidate failure mechanism, corresponding evidence mode and its early warning model path label, the verification criterion is used to judge the consistency of the quantitative evaluation result and the evidence mode, and the risk analysis specification is output; S3, based on the risk analysis specification, the task execution plan object is generated by the risk analysis task scheduling intelligent agent, the task execution plan object contains the preferred model path, the candidate hypothesis mapping table and the directional re-planning rule, the directional re-planning rule stipulates that when the verification criterion is not satisfied, the model path is switched according to the candidate hypothesis mapping table and the query specification is kept unchanged, so that the evaluation and explanation converge around the risk hypothesis, and the task execution plan object is output; S4, according to the query specification and the task execution plan object, generating a structured query instruction and obtaining a target data object, keeping the field and time boundary consistent under the constraint of the directional re-planning rule; S5, based on the target data object, the quantitative evaluation result is calculated according to the preferred model path, and the verification criterion is judged, the determined hypothesis identifier is generated when the verification criterion is satisfied, and the candidate hypothesis identifier is generated when the verification criterion is not satisfied; S6, based on the hypothesis identifier and the quantitative evaluation result, the knowledge base is retrieved with the hypothesis identifier as the anchor point, and the explanation and disposal report object is generated; S7, output the explanation and disposal report object as the major hazard source early warning explanation result.

2. The risk analysis method based on large model and early warning model cooperation according to claim 1, characterized in that, Step S1 is specifically: Taking the natural language task request as input, performing text normalization and sentence segmentation, establishing an analysis channel for dangerous chemical production device identifier, time range and monitoring index, eliminating irrelevant modifiers and redundant descriptions; Identify the dangerous chemical production device identifier from the natural language task request, merge the references and aliases, and determine a single and referable dangerous chemical production device identifier; Parse the time range, identify the start time and end time, unify the absolute time and relative time, and form a time range containing the start time and end time; Parse the monitoring index, map it to the index field, unify the unit of measurement and sampling interval, and assemble the analysis target description based on the determined dangerous chemical production device identifier, time range and monitoring index; Format check the analysis target description to ensure that the dangerous chemical production device identifier is unique, the time range contains the start time and end time, the monitoring index corresponds to the index field and the units are consistent, and the analysis target description can be directly referenced.

3. The risk analysis method based on large model and early warning model cooperation according to claim 1, characterized in that, Step S2 is specifically: Taking the analysis target description as input, calling the field instruction fine-tuning large language model to generate the risk analysis specification in three sections, clearly containing the query specification, the risk hypothesis set and the verification criterion, and establishing the reference relationship among the query specification, the risk hypothesis set and the verification criterion; According to the device identifier, time range and monitoring index in the analysis target description, a query specification is generated, the device identifier, time range and index field are limited, and item-by-item mapping and unification are performed, so that the query specification can be directly used for structured query instructions; A risk hypothesis set is generated, a candidate failure mechanism is taken as a unit, a corresponding evidence mode is determined, and a pre-warning model path label is bound for each candidate failure mechanism, so that the candidate failure mechanism, the evidence mode and the pre-warning model path label form a parameterized reference set; Verification criteria are generated, the evidence mode in the risk hypothesis set is used as a basis to set a judgment boundary and a comparison relationship, a quantitative evaluation result and a consistency determination method of the evidence mode are specified, and a consistent and inconsistent determination output is given; The consistency of the risk analysis specification is checked, the terminology consistency and reference closure of the query specification, the risk hypothesis set and the verification criteria are checked, and the constraint relationship between the pre-warning model path label and the device identifier, the time range and the index field is confirmed; The pre-warning model path label and the verification criteria are introduced in the risk analysis specification generation stage, the above constraints are taken as the input starting point of the subsequent task execution plan object, and it is ensured that the task arrangement and interpretation converge around the candidate failure mechanism; The query specification, the risk hypothesis set and the verification criteria are solidified in a unified format, the boundary consistency of the device identifier, the time range and the index field is maintained, and the risk analysis specification is output.

4. The risk analysis method based on large model and early warning model cooperation of claim 1, characterized in that, Step S3 is specifically: The risk analysis specification is taken as an input, the query specification, the risk hypothesis set and the verification criteria are parsed by the risk analysis task arrangement intelligent agent, the device identifier, the time range and the index field in the query specification are extracted, the candidate failure mechanism, the evidence mode and the pre-warning model path label in the risk hypothesis set are extracted, the judgment boundary and the comparison relationship in the verification criteria are extracted, and the field mapping relationship between the above contents and the task execution plan object is established; Based on the pre-warning model path label in the query specification and the risk hypothesis set, a preferred model path is determined, the pre-warning model path label with the same index field set as the required index field set in the query specification is preferentially selected, if there is no such pre-warning model path label, the pre-warning model path label with the largest number of required index fields is selected, and the binding mode of the device identifier, the time range and the index field is solidified in the preferred model path; A candidate hypothesis mapping table is constructed, the verification criteria are decomposed into judgment items according to the index field, each judgment item is specified with a candidate failure mechanism and its pre-warning model path label, the evidence mode corresponding to the specified pre-warning model path label is required to directly involve the index field corresponding to the judgment item, and the switching direction of the candidate failure mechanism and the preferred model path is recorded; Directional re-planning rules are generated, the judgment results of the verification criteria are taken as trigger conditions, when any judgment item does not meet the requirements, the corresponding pre-warning model path label is switched according to the candidate hypothesis mapping table, the query specification is kept unchanged, the device identifier, the time range and the index field are prohibited from being modified, and the judgment boundary and the comparison relationship of the verification criteria are kept unchanged; The switching sequence control is set in the directional re-planning rule, the switching is triggered one by one according to the listed order of the judgment items in the verification criterion, parallel triggering is prohibited, and the output definition of the quantitative evaluation result in the task execution plan object is used after the switching; The assumption identification generation rule is preset in the task execution plan object, the candidate failure mechanism is corresponded to the early warning model path label to form the assumption identification, and the pointing of the assumption identification to be generated is updated when the path is switched, so that the assumption identification can be consistent with the query specification; The preferred model path, the candidate assumption mapping table and the directional re-planning rule are assembled into the task execution plan object, and the task execution plan object is output.

5. The risk analysis method based on large model and early warning model cooperation of claim 1, characterized in that, Step S4 is specifically: The query specification and the task execution plan object are taken as inputs, the device identification, the time range and the index field in the query specification are parsed, the preferred model path and the directional re-planning rule in the task execution plan object are parsed, and the binding mode in the task execution plan object is used as a constraint for generating a structured query instruction; According to the query specification, a structured query instruction is generated, the device identification is limited to the device identification in the query specification, the time boundary is limited to the time range in the query specification, and the field is limited to the index field in the query specification, to form a structured query instruction that can be directly executed; The structured query instruction is executed to obtain a target data object, when the directional re-planning rule triggers path switching, the query specification is kept unchanged according to the directional re-planning rule, the structured query instruction is repeatedly executed, and it is ensured that the device identification, the time range and the index field are consistent with the query specification; The target data object is subjected to consistency checking, the index field of the target data object is checked to cover the index field in the query specification, the time boundary of the target data object is checked to cover the time range in the query specification, and it is confirmed that the target data object can be directly referenced by the preferred model path or the early warning model path after switching.

6. The risk analysis method based on large model and early warning model cooperation of claim 1, characterized in that, Step S5 is specifically: The target data object and the preferred model path are taken as inputs, the target data object is subjected to quantitative calculation according to the preferred model path, and a quantitative evaluation result is generated; According to the judgment boundary and the contrast relationship in the verification criterion, the quantitative evaluation result is subjected to consistency judgment, and a judgment result that meets or does not meet is obtained; When the consistency judgment meets, the candidate failure mechanism and the early warning model path label are combined to generate a determined assumption identification, and the determined assumption identification is kept consistent with the device identification, the time range and the index field; When the consistency judgment does not meet, the candidate failure mechanism and the early warning model path label are combined to generate a candidate assumption identification, and the candidate assumption identification is kept consistent with the device identification, the time range and the index field.

7. The risk analysis method based on large model and early warning model cooperation of claim 1, characterized in that, Step S6 is specifically: The assumption identification and the quantitative evaluation result are taken as inputs, the knowledge base is retrieved with the assumption identification as an anchor point, the retrieval range is limited to the candidate failure mechanism and the evidence mode corresponding to the assumption identification, and the boundaries of the device identification, the time range and the index field in the query specification are used; According to the evidence mode corresponding to the assumption identification in the knowledge base, the quantitative evaluation result is contrasted by using the judgment boundary and the contrast relationship in the verification criterion, to form a contrast conclusion of mechanism consistency or inconsistency; According to the comparison conclusion, the correspondence between the hypothesis identifier, the quantitative evaluation result and the evidence mode is written into the explanation and treatment report object, and the correspondence with the device identifier, the time range and the index field in the query specification is clearly indicated in the explanation and treatment report object; According to the correspondence between the knowledge base and the hypothesis identifier, the comparison conclusion is merged to generate the explanation and treatment report object, and the device identifier, the time range and the index field in the query specification are marked, and the explanation and treatment report object is output.

8. The risk analysis method based on large model and early warning model cooperation of claim 1, characterized in that, Step S7 is specifically: Aligning the explanation and treatment report object with the query specification, marking the device identifier, the time range and the index field in the query specification, taking the hypothesis identifier and the quantitative evaluation result as the reference items of the explanation and treatment report object, and forming the corresponding explanation and treatment content based on the determination conclusion of the verification criterion, listing the candidate failure mechanism and the early warning model path label corresponding to the hypothesis identifier in the explanation and treatment report object; The consistency of the explanation and treatment report object is checked, and it is confirmed that the hypothesis identifier, the quantitative evaluation result and the device identifier, the time range and the index field in the query specification remain consistent, it is confirmed that the device identifier, the time range and the index field do not change in the two cases of triggering and not triggering the directional re-planning rule, and the consistency checking conclusion is retained; In the explanation and treatment report object, the determined hypothesis identifier or the candidate hypothesis identifier is marked, and the early warning model path label and the determination conclusion of the verification criterion are associated, so that each item of content can be directly traced by the query specification, while keeping the consistent reference of the device identifier, the time range and the index field, and the explanation and treatment report object is output as the major hazard source early warning explanation result.

9. A risk analysis agent system based on the cooperation of large models and early warning models, used to execute the risk analysis method based on the cooperation of large models and early warning models according to any one of claims 1 to 8, comprising: A natural language analysis module for analyzing the dangerous chemical production device identifier, the time range and the monitoring index, and generating an analysis target description; A risk analysis specification generation module for calling a domain instruction fine-tuned large language model to generate a query specification, a risk hypothesis set and a verification criterion; A risk analysis task scheduling module for generating a task execution plan object containing a preferred model path, a candidate hypothesis mapping table and a directional re-planning rule; A query and data acquisition module for generating a structured query instruction according to the query specification and the task execution plan object and acquiring a target data object; A quantitative evaluation and hypothesis identification module for calculating quantitative evaluation results according to the preferred model path and generating determined hypothesis identifiers or candidate hypothesis identifiers according to the verification criterion; A knowledge retrieval and explanation generation module for retrieving a knowledge base with the hypothesis identifier as an anchor point and generating an explanation and treatment report object according to the evidence mode and the quantitative evaluation result; An alignment and output module for aligning the explanation and treatment report object with the query specification and performing consistency checking, associating the early warning model path label with the determination conclusion, and outputting the major hazard source early warning explanation result.

Images