Integrated Knowledge-Graph-Based Chemical Safety Early Warning System Including AI Extraction-Inference Output Governance Layer

Abstract

The present invention relates to a knowledge graph-based chemical safety early warning integrated system comprising a governance layer of AI extraction and inference results. The system comprises eight components: (a) an integrated ontology unit, (b) an AI extraction and inference unit, (c) a KG generation unit, (d) a mixed risk calculation unit, (e) a distributed learning unit, (f) an HVD curation unit, (g) an AI control unit, and (h) an alarm issuance unit. (b) The AI ​​extraction and inference unit is implemented as at least one of a pre-trained NER model, a large-scale language model (LLM), and an Agent AI system. (g) The AI ​​control unit calculates an integrated uncertainty score for the output of (b) the AI ​​extraction and inference unit by combining SHACL integrity verification, regulatory compliance verification, triple identifier mapping agreement verification, and hallucination detection, escalates results exceeding a threshold to a human reviewer, and feeds the review results back to (b) the AI ​​extraction and inference unit as an active learning signal. The present invention ensures that the safety management output quality of the integrated system is guaranteed to be above a certain level even in future AI technology evolution (NER → LLM → Agent → next-generation AI) by (b) regardless of the modality type of the AI ​​extraction and inference unit, and (g) the AI ​​control unit guarantees output quality with the same control protocol. A closed-loop structure is formed with three feedback loops of distributed learning, HVD curation, and the AI ​​control unit, thereby achieving an improvement in precision over time, and the present invention protects the synergy of the applicant's six separate applications as a single system claim.

Description

Technology Field

[0001] The present invention relates to an integrated monitoring system in the field of chemical safety, and more specifically, to a system that provides integrated chemical safety early warning by combining an integrated ontology, an AI extraction and inference unit that extracts or infers five entities from a Korean accident report, six major components including the generation of a spatiotemporal causal knowledge graph, the calculation of the risk of exposure to a mixture of complex chemical substances, distributed federated learning at the workplace level, and high-value data (HVD) curation, an AI control unit that verifies and controls the output results of the AI ​​extraction and inference unit based on SHACL integrity, regulatory compliance, and triple identifier match, and an alarm issuance unit.

[0002] In particular, the present invention relates to an integrated system protected by a single system claim by integrally combining eight modules, and to a chemical safety early warning system that improves accuracy over time and adapts to workplace characteristics through data flow and feedback structures between each component. Furthermore, the present invention relates to a structure in which the extraction modality is generalized so that an AI engine for extracting or inferring five entities from a Korean accident report can adopt any one of a pre-trained NER model, a Large Language Model (hereinafter referred to as “LLM”), or an Agent AI system.

[0003] Furthermore, the essence of the present invention lies not in the type of the AI ​​extraction and inference unit itself, but in the AI ​​Governance layer that automatically verifies, filters, and feeds back the results produced by the AI ​​extraction and inference unit based on SHACL integrity constraints, compliance with the three major regulations, triple identifier mapping consistency, and hallucination detection signals. Accordingly, the invention relates to a technology-neutral integrated system in which the quality of the safety management output of the integrated system is guaranteed to be above a certain level, regardless of what type of AI model is adopted in the AI ​​extraction and inference unit. Background Technology

[0005] Safety management at workplaces handling chemical substances is critical for protecting workers' health, safeguarding the environment, and preventing business losses; consequently, compliance with the obligation to conduct risk assessments under Article 36 and the obligation to implement workplace safety measures under Article 63 of the Korean Occupational Safety and Health Act is required. However, conventional chemical safety monitoring systems have limitations as they are restricted to single-function modules and fail to provide integrated safety management.

[0006] The accident investigation report system of the U.S. Chemical Safety Board (CSB) is known as Prior Art 1. Prior Art 1 has limitations in that it focuses on post-accident investigation and report writing, and therefore cannot provide accident prevention and real-time early warning functions.

[0007] Prior art 2 is known as a single-site monitoring system based on SCADA (Supervisory Control and Data Acquisition) or PLC (Programmable Logic Controller). Prior art 2 is limited to the collection of measurement data and simple threshold comparison for a single site, and has limitations in that it does not provide integrated functions such as inferring causal interactions between chemicals, calculating synergistic effects of combined exposures, and adaptive learning for each site.

[0008] Prior art 3 is known as an IoT sensor-based leak detection system (e.g., a gas detector). Prior art 3 is limited to a simple detection function that issues an alarm when a critical concentration of a specific substance is exceeded, and has limitations in that it does not provide the integrated safety management functions provided by the present invention, such as complex exposure synergy, spatiotemporal causal inference, and workplace adaptive learning.

[0009] Prior Art 4 is known as an LLM or Agent AI-based automatic natural language accident report analysis system that is being introduced in the field of chemical safety. Prior Art 4 utilizes the extensive prior learning knowledge of LLM or Agent AI to extract or infer information such as chemical substances, facilities, accident types, damage scale, and response measures from Korean accident reports. However, Prior Art 4 has limitations in that it fails to guarantee the safety management suitability of the extraction and inference results because it does not incorporate a control layer that automatically detects and filters issues such as hallucination inherent to LLM and Agent AI, inaccurate chemical name inference, and facility classifications that differ from the facts. In other words, while Prior Art 4 shows advancements in the extraction capabilities of the AI ​​model itself, it lacks a governance structure that automates the verification, normalization, and feedback of the AI ​​output.

[0010] In addition, the applicant has filed separate Korean patent applications for each of the six components constituting the integrated system of the present invention. However, since separate applications for each component are limited to individual functions and do not protect the synergistic effects of the integrated system, a separate claim for the integrated system is required to protect the data flow and feedback structure itself, including the integrated combination of the six components and the AI ​​control layer.

[0011] Ultimately, conventional technology has the following limitations. First, it is limited to post-investigation or simple detection and fails to provide integrated early warning functions. Second, since only a single module is operated, it lacks integrated functions such as chemical causal inference, synergy calculation, and workplace adaptation. Third, the absence of a feedback structure with external KGs makes it impossible to improve accuracy over time. Fourth, integrated operation across multiple workplaces is difficult because a balance between the protection of workplace trade secrets and learning contributions is not achieved. Fifth, even when adopting LLM / Agent AI, it is unsuitable for direct application to chemical safety decision-making due to the lack of a layer to automatically detect and control hallucinations in AI output, regulatory violations, and identifier mismapping.

[0012] Therefore, a new chemical safety early warning integrated system is required that integrates eight major components: an integrated ontology, an AI extraction and inference unit (at least one of NER model, LLM, and Agent AI), a spatiotemporal causal KG, a mixed hazard calculation, distributed federated learning, HVD curation, an AI control unit, and an alarm issuance unit. Prior art literature

[0014] (Prior Art 1) US Chemical Safety and Hazard Investigation Board, “Accident Investigation Report and Recommendations” https: / / www.csb.gov / investigations / , accessed 2024. (Prior Art 2) Stouffer, K. et al., “to Industrial Control Systems (ICS) Security”, NIST Special Publication 800-82 Revision 3, 2023. (Prior Art 3) Aleixandre-Tudo, JL et al., “IoT-Based Real-Time Gas Leak Detection: A Review”, Sensors, 23(4): 2018-2042, 2023. (Prior Art 4) Bran AM et al., “Augmenting large-language-models with chemistry tools”, Nature Machine Intelligence, 5: 525-535, 2024 (Representative prior art example of AI-based LLM / Agent analysis in the chemical field). The problem to be solved

[0015] The present invention was devised to solve the problems of the prior art as described above, and aims to solve the following problems.

[0016] First, an integrated system is provided that provides chemical safety early warning by integrally combining eight major components: an integrated ontology, an AI extraction and inference unit (at least one of a NER model, LLM, and Agent AI), a spatiotemporal causal KG, mixed risk calculation, distributed federated learning, HVD curation, an AI control unit, and an alarm issuance unit.

[0017] Second, a technology-neutral integrated system is provided by including an AI control unit as an explicit component within the integrated system that automatically detects, filters, and feeds back hallucinations, regulatory violations, and identifier mismappings in the AI ​​output regardless of the modality of the AI ​​extraction and inference unit (i.e., whether NER model, LLM, or Agent AI is adopted), thereby ensuring the safety management output quality of the integrated system at a level above a certain standard regardless of the type of AI model.

[0018] Third, through the data flow and feedback structure between the eight major components (distributed learning → KG feedback, HVD → KG feedback, AI control unit → AI extraction / inference unit feedback), an integrated system with a closed-loop structure is provided in which precision improves over time.

[0019] Fourth, it provides an alarm issuance unit that combines accident scenario matching and 3-stage alarm (green / yellow / red) classification, thereby providing precise early warning compared to simple threshold comparison.

[0020] Fifth, the synergy of six separate applications is protected by a single system claim combined with an AI control unit, thereby securing rights that enable integrated commercialization and licensing.

[0021] Sixth, it provides a flexible integrated system that can be utilized to fulfill the obligations under Articles 36 and 63 of the Industrial Safety and Health Act and can be operated as a single workplace or a consortium of multiple workplaces. means of solving the problem

[0023] To solve the above problem, the present invention provides a knowledge graph-based chemical safety early warning integrated system. The integrated system comprises eight components: (a) an integrated ontology unit, (b) an AI extraction and inference unit, (c) a KG generation unit, (d) a mixed risk calculation unit, (e) a distributed learning unit, (f) an HVD curation unit, (g) an AI control unit, and (h) an alarm issuance unit.

[0024] (a) The integrated ontology part defines six object classes (Chemical, Accident, HealthImpact, Workplace, Worker, Environment) and four relationship classes (causedBy, leadsToImpact, locatedIn, exposedTo) of the chemical safety domain. In one embodiment of the present invention, the chemical safety integrated ontology schema of the applicant’s separate Korean patent application (hereinafter referred to as the 'A-2 system') may be utilized, but the present invention is applicable to any chemical safety ontology and is not limited thereto.

[0025] (b) The AI ​​extraction and inference unit extracts or infers five entities—chemical name, accident type, time, place, and symptom—from the natural language text of a Korean accident report and converts them into an RDF triple (subject, predicate, object) format. The AI ​​extraction and inference unit comprises at least one of (b-1) a pre-trained Korean Named Entity Recognition (NER) model, (b-2) a pre-trained Large Language Model (LLM) that operates on a prompt basis, and (b-3) an Agent AI system having a tool calling function. In one embodiment of the present invention, the implementation based on (b-1) the NER model may utilize the Korean accident report Named Entity Recognition System (hereinafter referred to as the 'B-2 System') of the applicant's separate Korean patent application, but the present invention is applicable to any one of the NER model, LLM, and Agent AI, or a combination thereof, and is not limited thereto.

[0026] (c) The KG generation unit generates a spatiotemporal causal knowledge graph by assigning spatiotemporal first-degree attributes (τ time delay, W causal strength, β spatial decay coefficient, C context vector) to an RDF triple input from the AI ​​extraction and inference unit. In one embodiment of the present invention, the spatiotemporal causal knowledge graph system of the applicant’s separate Korean patent application (hereinafter referred to as the 'B-1 system') may be utilized, but the present invention is applicable to any spatiotemporal causal knowledge graph and is not limited thereto.

[0027] (d) The mixed risk calculation unit extracts causal pathways from a set of chemical substances and KG using SPARQL queries and calculates pathway-weighted synergies to calculate a composite risk score R = HI + α × S_total. In one embodiment of the present invention, the knowledge graph pathway-based complex chemical substance mixed exposure risk prediction system (hereinafter referred to as the 'B-3 system') of the applicant's separate Korean patent application may be utilized, but the present invention may apply any means of calculating synergies and is not limited thereto.

[0028] (e) The distributed learning unit performs distributed federated learning with differential ε-privacy based on the workplace trade secret (CBI) grade to extract workplace adaptation parameters (α, θ1 / θ2, λ, w_i). In one embodiment of the present invention, the workplace unit distributed federated learning system of the applicant’s separate Korean patent application (hereinafter referred to as the 'B-4 system') may be utilized, but the present invention is applicable to any distributed learning system and is not limited thereto.

[0029] (f) The HVD curation department evaluates five quality indicators (completeness, accuracy, timeliness, consistency, accessibility) for external data candidates (new chemical substances, public datasets, etc.) to automatically curate high-value data (HVD) and register it in the KG. In one embodiment of the present invention, the chemical safety HVD automatic curation system of the applicant’s separate Korean patent application (hereinafter referred to as the 'B-5 system') may be utilized, but the present invention may apply any curation means and is not limited thereto.

[0030] (g) The AI ​​control unit performs the following functions regarding the output results of the above (b) AI extraction and inference unit: (g-1) verification of consistency based on SHACL integrity constraints, (g-2) verification of legal consistency based on the statutory classification of the Chemical Substances Control Act, the Occupational Safety and Health Act, and the Dangerous Goods Safety Management Act, (g-3) verification of mapping consistency between CAS, GHS, and KOSHA triple identifiers, (g-4) calculation of uncertainty scores combined with AI hallucination detection signals and comparison of thresholds, and (g-5) automatic escalation of uncertainty outputs above the threshold to a human reviewer queue and feed back the review results as active learning signals to the (b) AI extraction and inference unit. The AI ​​control unit functions as a cross-cutting module of the integrated system in that it verifies, filters, and feeds back the output using the same control protocol regardless of whether the (b) AI extraction and inference unit is implemented as a NER model, LLM, or Agent AI.

[0031] (h) The alarm issuing unit combines the composite risk score R of the mixed risk calculation unit with the accident scenario DB matching result to issue a 3-stage alarm (green / yellow / red) and provides integrated safety management output.

[0032] In addition, the present invention provides a method for operating the integrated system. The method includes (1) a data input step, (2) an ontology mapping and AI extraction / inference step, (3) an output verification, filtering, and feedback step of the AI ​​control unit, (4) a KG generation step, (5) a mixed risk calculation step, (6) a distributed learning and HVD curation feedback step, and (7) an alarm issuance step. Effects of the invention

[0034] A knowledge graph-based chemical safety early warning integrated system including a control layer for AI extraction and inference results according to the present invention has the following effects.

[0035] First, comprehensive chemical safety early warning is possible compared to conventional single-module systems (SCADA, IoT, etc.) through the integrated combination of eight components (integrated ontology, AI extraction / inference unit, KG generation, mixed risk calculation, distributed learning, HVD curation, AI control unit, and alarm issuance).

[0036] Second, (b) regardless of whether the AI ​​extraction and inference unit is implemented as a NER model, LLM, or Agent AI, (g) the AI ​​control unit verifies, filters, and recirculates the output using the same control protocol, so the quality of the safety management output of the integrated system is guaranteed to be above a certain level regardless of the type of AI model and future evolution of AI technology. In other words, the present invention provides an AI modality-neutral and future-oriented integrated chemical safety system.

[0037] Third, through the KG feedback structure of the distributed learning unit and HVD curation unit and the active learning feedback structure (closed loop) of the AI ​​control unit, accuracy is continuously improved over time, and the listing of new chemical substances and workplace adaptation are performed automatically.

[0038] Fourth, through accident scenario matching and three-stage alarm classification, precise early warning is possible compared to simple threshold comparison, and alarm fatigue is alleviated, ensuring a rapid response in emergency situations.

[0039] Fifth, by protecting the synergy of six separate applications through a single system claim combined with the AI ​​control unit, the present invention possesses independent legal value that does not overlap with the separate applications of individual modules. In particular, even if the AI ​​extraction and inference unit is bypassed or replaced by LLM or Agent AI, utilizing the AI ​​control unit ensures that it is encompassed by the rights of the present invention, making it difficult to evade the scope of rights.

[0040] Sixth, it can be directly utilized for the implementation of risk assessment and workplace safety measures under the Industrial Safety and Health Act, and can be flexibly operated as a single workplace or a consortium of multiple workplaces, making it applicable to various industrial sectors such as oil refining / petrochemical complexes, semiconductor cleanrooms, pharmaceutical / life science facilities, medical institutions, and environmental management corporations.

[0041] Seventh, the standardization of the integrated system's SPARQL / REST API facilitates interoperability with external systems and possesses industrial infrastructure value that makes it suitable for adoption as an integrated monitoring system by the Korea Occupational Safety and Health Agency or the Ministry of Environment. Brief explanation of the drawing

[0043] FIG. 1 is a diagram showing the overall structure of a knowledge graph-based chemical safety early warning integrated system according to one embodiment of the present invention. FIG. 2 is a diagram showing the detailed configuration of (a) an integrated ontology unit according to one embodiment of the present invention. FIG. 3 is a diagram showing the detailed configuration of (b) an AI extraction and inference unit according to one embodiment of the present invention, illustrating a modality generalization structure implemented as at least one of a NER model, an LLM, and an Agent AI. FIG. 4 is a diagram showing the detailed configuration of (c) a KG generation unit according to one embodiment of the present invention. FIG. 5 is a diagram showing the detailed configuration of (d) a mixed risk calculation unit according to one embodiment of the present invention. FIG. 6 is a diagram showing the detailed configuration of (e) a distributed learning unit according to one embodiment of the present invention. FIG. 7 is a diagram showing the detailed configuration of (f) HVD curation unit according to one embodiment of the present invention. FIG. 8 is a diagram showing the detailed configuration of (g) an AI control unit according to one embodiment of the present invention, illustrating five major control procedures for the output of the AI ​​extraction and inference unit: SHACL verification, legal compliance verification, triple identifier matching verification, hallucination detection, and active learning feedback. FIG. 9 is a drawing showing the detailed configuration of (h) an alarm issuing unit according to one embodiment of the present invention. FIG. 10 is a diagram showing data flow and closed-loop feedback structure within an integrated system according to an embodiment of the present invention (distributed learning → KG feedback, HVD → feedback, AI control unit → AI extraction / inference unit feedback). FIG. 11 is a diagram showing the results of the input / output comparison by modality (NER model, LLM, Agent AI) of the AI ​​extraction and inference unit and the integrated verification of the AI ​​control unit according to one embodiment of the present invention. Specific details for implementing the invention

[0044] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings. In describing the present invention, if it is determined that a detailed description of related known functions or configurations may obscure the essence of the present invention, such detailed description will be omitted. The terms used in this specification are for describing embodiments of the present invention and are not intended to limit the scope of the present invention.

[0045] In this specification, the term “AI extraction and inference unit” refers to a module that extracts or infers five entities—chemical name, accident type, time, place, and symptom—from natural language text of a Korean accident report and converts them into an RDF triple format, and is a higher-level concept that can be implemented as at least one of a pre-trained Korean NER model, a pre-trained LLM, or an Agent AI system having a tool call function.

[0046] In this specification, “model” means a Korean named entity recognition model pre-trained by BIO tagging or sequence labeling methods, and includes a model that combines a pre-trained encoder such as KoBERT, KLUE-RoBERTa, KoELECTRA, and domain-adaptive additional training.

[0047] In this specification, “LLM (Large Language Model)” refers to a large-scale pre-trained language model with a transformer structure that performs prompt-based extraction; for example, a Korean language-supported LLM disclosed as of the filing date of this application falls under this category, but the present invention is not limited to a specific model.

[0048] In this specification, the term “Agent AI system” refers to an AI system in which an LLM or an equivalent inference engine is equipped with a tool calling function that calls external tools (search, database query, calculator, RDF converter, etc.) and autonomously performs multi-stage inference.

[0049] In this specification, “AI Governance” refers to a set of control procedures that automatically verify, filter, and provide feedback on the results produced by the AI ​​extraction and inference unit by applying SHACL integrity constraints, legal compliance, triple identifier mapping consistency, and hallucination detection signals. AI Governance is applied using the same protocol regardless of the modality of the AI ​​extraction and inference unit.

[0050] In this specification, “hallucination” refers to a phenomenon in which an AI extraction or inference unit, such as an LLM or Agent AI, generates output that is unfounded in the input text or inconsistent with the facts.

[0051] In this specification, the term “closed-loop feedback structure” refers to a cyclic structure in which the output result of an integrated system is input back into a KG or AI extraction / inference unit and utilized for updating a model or data. The present invention includes three loops: distributed learning unit → KG feedback, HVD curation unit → KG feedback, and AI control unit → AI extraction / inference unit feedback.

[0052] Example 1 — Overall structure of the integrated system

[0053] FIG. 1 is a diagram showing the overall structure of a knowledge graph-based chemical safety early warning integrated system according to an embodiment of the present invention. Referring to FIG. 1, the integrated system according to an embodiment of the present invention includes a data input unit (10), (a) an integrated ontology unit (100), (b) an AI extraction and inference unit (200), (c) a KG generation unit (300), (d) a mixed risk calculation unit (400), (e) a distributed learning unit (500), (f) an HVD curation unit (600), (g) an AI control unit (750), and (h) an alarm issuance unit (700), and an integrated safety management output unit (800).

[0054] The data input unit (10) receives accident reports, measurement data, chemical specifications, and environmental condition data and transmits them to the subsequent configuration unit.

[0055] (a) The integrated ontology unit (100) provides an integrated ontology that defines six object classes (Chemical, Accident, HealthImpact, Workplace, Worker, Environment) and four major relationship classes (causedBy, leadsToImpact, locatedIn, exposedTo) of the chemical safety domain. The integrated ontology is used as a standard for semantic verification in the subsequent AI extraction / inference unit (200), AI control unit (750), and KG generation unit (300). In one embodiment of the present invention, the integrated ontology may utilize an ontology defined by the system A-2 of the applicant's separate Korean patent application, but the present invention is applicable to any chemical safety ontology and is not limited thereto.

[0056] (b) The AI ​​extraction and inference unit (200) converts the natural language text of a Korean thought report input from the data input unit (10) into an RDF triple according to the class of the integrated ontology. In one embodiment of the present invention, the AI ​​extraction and inference unit (200) is implemented as at least one of (b-1) a pre-trained NER model, (b-2) an LLM, and (b-3) an Agent AI system. The detailed structure thereof will be described later with reference to FIG. 3. (b-1) The NER model-based implementation may utilize the B-2 system of the applicant’s separate Korean patent application, but the present invention is applicable to any one of the NER model, LLM, and Agent AI, or a combination thereof, and is not limited thereto.

[0057] (c) The KG generation unit (300) generates a spatiotemporal causal knowledge graph by assigning spatiotemporal first-degree attributes (τ time delay, W causal strength, β spatial damping coefficient, C context vector) to RDF triples that have passed verification by the AI ​​control unit (750). In one embodiment of the present invention, the KG generation unit (300) may utilize the B-1 system of the applicant’s separate Korean patent application, but the present invention is applicable to any spatiotemporal causal knowledge graph and is not limited thereto.

[0058] (d) The mixed risk calculation unit (400) extracts causal pathways from a chemical set and KG using SPARQL queries and calculates pathway-weighted synergies to calculate a composite risk score R = HI + α × S_total. In one embodiment of the present invention, the mixed risk calculation unit (400) may utilize the B-3 system of the applicant’s separate Korean patent application, but the present invention may apply any synergy calculation means and is not limited thereto.

[0059] (e) The distributed learning unit (500) performs differential ε-variance federated learning according to the business secret grade of the workplace to extract the workplace adaptation parameters α, θ1 / θ2, λ, and w_i, and feeds back the Bayesian updated causal strength to the KG generation unit (300). In one embodiment of the present invention, the distributed learning unit (500) may utilize the B-4 system of the applicant's separate Korean patent application, but the present invention is applicable to any distributed learning system and is not limited thereto.

[0060] (f) The HVD curation unit (600) evaluates external data candidates using five quality indicators to automatically curate the HVD and feeds new data back to the KG generation unit (300). In one embodiment of the present invention, the HVD curation unit (600) may utilize the B-5 system of the applicant’s separate Korean patent application, but the present invention may apply any curation means and is not limited thereto.

[0061] (g) The AI ​​control unit (750) receives the output results of (b) the AI ​​extraction and inference unit (200), performs SHACL integrity verification, legal compliance verification, triple identifier mapping match verification, calculation of hallucination detection signal combination uncertainty score, and threshold comparison, automatically escalates uncertainty outputs above the threshold to a human reviewer queue, and feeds the review results back to (b) the AI ​​extraction and inference unit (200) as active learning signals. The detailed structure of the AI ​​control unit (750) will be described later with reference to FIG. 8.

[0062] (h) The alarm issuing unit (700) combines the composite risk score R of the mixed risk calculation unit (400) with the accident scenario DB matching result to issue a 3-stage alarm (green / yellow / red) and transmits the output to the integrated safety management output unit (800).

[0063] The integrated safety management output unit (800) is linked with external systems through SPARQL API and REST API and can be operated as a single workplace or a consortium of multiple workplaces.

[0064] Example 2 ― Modality Generalization Structure of AI Extraction and Inference Unit

[0065] FIG. 3 is a diagram showing the detailed configuration of (b) an AI extraction and inference unit according to one embodiment of the present invention. Referring to FIG. 3, the AI ​​extraction and inference unit (200) according to one embodiment of the present invention includes at least one of (b-1) a NER model modality (210), (b-2) an LLM modality (220), and (b-3) an Agent AI modality (230), and includes an extraction and inference result integration unit (240).

[0066] (b-1) The NER modality (210) is implemented as a model that simultaneously performs BIO tagging for 5 entities, including chemical substances, facilities, accident types, damage scale, and response measures, through a Korean dictionary-learned encoder (KoBERT, KLUE-RoBERTa, KoELECTRA, etc.) and a domain-adapted transformer backbone. For each token of an input sentence, it outputs the probability distribution of BIO tags for each of the 5 entities and ensures sequence consistency through CRF decoding.

[0067] (b-2) The LLM modality (220) is implemented by inputting an extraction prompt template defined in the present invention into a Korean-supported pre-learned LLM to receive 5 entities as JSON or structured text output. The extraction prompt includes class definitions of the integrated ontology (100) and name and format specifications of the 5 entities, and may utilize a structured output or function calling interface for format consistency of the LLM output.

[0068] (b-3) The Agent AI modality (230) is implemented in such a way that an LLM-based inference engine performs multi-stage inference by calling external tools (e.g., Korean morphological analyzer, CAS·GHS·KOSHA identifier lookup tool, chemical name normalizer, hazardous material type classifier, etc.). The Agent AI modality (230) can demonstrate superior extraction performance compared to the NER modality or simple LLM modality in complex sentences or multiple accident reports that are difficult to extract, but at the same time, the risk of hallucination increases, so verification by the AI ​​control unit (750) is essential.

[0069] The extraction and inference result integration unit (240) integrates the 5 entities output from one or more of the modalities (b-1) to (b-3) into a single result set (CHEM, FACI, ACCD, DAMG, RESP) and transmits it to the AI ​​control unit (750). When two or more modalities are operated simultaneously, the consistency of the results between the modalities is used as one of the hallucination detection signals of the AI ​​control unit.

[0070] In one embodiment of the present invention, the selection of modalities of the AI ​​extraction and inference unit (200) may be determined as follows depending on the operating environment. First, in an environment where sufficient labeled Korean accident report training data is secured and the consistency and predictability of the extraction results are prioritized, (b-1) NER modality is adopted. Second, in an environment where labeled data is insufficient and generalization ability for accident reports of various formats is required, (b-2) LLM modality is adopted. Third, in an environment where multi-stage inference (e.g., accident report → chemical identification → CAS number lookup → GHS classification derivation) is required or external database linkage is required, (b-3) Agent AI modality is adopted. However, the specific selection of modalities does not affect the essence of the present invention, and the control protocol of the AI ​​control unit (750) is applied equally to all modalities.

[0071] Example 3 ― Detailed Structure and Control Procedure of the AI ​​Control Unit

[0072] FIG. 8 is a diagram showing the detailed configuration of (g) an AI control unit according to one embodiment of the present invention. Referring to FIG. 8, the AI ​​control unit (750) according to one embodiment of the present invention includes an input integration unit (751), (g-1) a SHACL integrity verification unit (752), (g-2) a rule consistency verification unit (753), (g-3) a triple identifier mapping match verification unit (754), (g-4) a hallucination detection and uncertainty score calculation unit (755), (g-5) a human reviewer escalation and active learning feedback unit (756) and a control result output unit (757).

[0073] The input integration unit (751) receives the output 5 entities (CHEM, FACI, ACCD, DAMG, RESP) of the (b) AI extraction and inference unit (200) and transmits them to the subsequent verification unit. Regardless of whether the AI ​​extraction and inference unit is implemented as a NER modality, an LLM modality, or an Agent AI modality, the input integration unit (751) integrates them into the same input format and proceeds with subsequent verification. Thus, the AI ​​control unit (750) operates as a modality-agnostic control protocol.

[0074] (g-1) The SHACL integrity verification unit (752) applies the SHACL Shape of the integrated ontology (100) to the input 5 entities to verify whether there is a violation of cardinality, datatype, value range, and referential integrity constraints. For example, it automatically verifies the “time order constraint of response-effect relationships (time of response action > time of accident)” and “facility hierarchy constraint of material-facility relationships.”

[0075] (g-2) The legal consistency verification unit (753) verifies the consistency of the extracted chemical substance, facility, and accident type entities according to the statutory classification system of the Chemical Substances Control Act, the Occupational Safety and Health Act, and the Dangerous Goods Safety Control Act. For example, it automatically verifies whether “the classification of hazardous chemical substances conforms to the attached table of the Enforcement Rule of the Chemical Substances Control Act” and “the accident type maps to one of the five accident type classifications of the Dangerous Goods Safety Control Act.” If the LLM or Agent AI modality outputs a legal classification that is different from the facts, this verification unit automatically detects it.

[0076] (g-3) The triple identifier mapping consistency verification unit (754) verifies the triple identifier mapping consistency of CAS number, GHS classification code, and KOSHA code for the extracted chemical substance. If the identifier generated by the AI ​​extraction and inference unit does not match the mapping table of an external standard dataset (e.g., PubChem, KOSHA MSDS database), the discrepancy is automatically detected. In particular, the verification unit automatically detects hallucinogenic outputs such as “generation of a non-existent CAS number” or “GHS classification that does not match the chemical name,” which frequently occur in LLM modalities.

[0077] (g-4) The hallucination detection and uncertainty score calculation unit (755) calculates an integrated uncertainty score u(x) by combining the verification results of (g-1) to (g-3). In one embodiment, u(x) is defined as a weighted sum of α·u_SHACL(x) + β·u_rule(x) + γ·u_identifier(x) + δ·u_modality(x), and u_modality(x) is calculated as the degree of output discrepancy between modalities when two or more modalities (NER·LLM·Agent) are operated simultaneously. α, β, γ, and δ are normalized as weights such that their sum is 1, and can be adjusted according to the operating environment, but are not limited thereto.

[0078] The calculated integrated uncertainty score u(x) is compared with a predefined threshold θ. In one embodiment of the present invention, θ can be determined within a range of 0.30 to 0.50. If the threshold θ is less than 0.30, almost all extraction results are escalated to human reviewers, resulting in an excessive burden on the reviewers; if the threshold exceeds 0.50, illusionary results are automatically passed through, introducing errors into the safety management output. Therefore, the range of 0.30 to 0.50 has critical significance as it represents a balance point between the burden on the reviewers and the quality of the output. However, the specific value of the threshold may be adjusted according to the operating environment and the distribution of AI modalities, but is not limited thereto.

[0079] (g-5) The human reviewer escalation and active learning feedback unit (756) automatically loads extraction results with an integrated uncertainty score greater than or equal to θ into a human reviewer queue and accumulates reviewer labeling results to feed back as an active learning signal to (b) the AI ​​extraction and inference unit (200). (b-1) In the case of the NER modality, it is used for model retraining, and (b-2) In the case of the LLM modality and (b-3) Agent AI modality, it is used for updating few-shot examples of extraction prompts or updating tool call policies. This feedback ensures consistency of the control protocol regardless of the type of AI modality.

[0080] The control result output unit (757) transmits the extraction result (i.e., control-passed result) with an integrated uncertainty score of less than θ to the KG generation unit (300). As a result, only verified RDF triples are input to the steps after the KG generation unit, and hallucinogenic outputs are prevented from being input into the KG or risk score calculation.

[0081] The AI ​​control unit (750) of the present invention has the following distinguishing features. First, the AI ​​control unit (750) verifies, filters, and feeds back the output using the same control protocol regardless of the modality type of (b) the AI ​​extraction and inference unit (200). Second, the AI ​​control unit (750) complements the limitations of a single verification standard by combining triple verification of SHACL integrity, legal compliance, and triple identifier matching. Third, the AI ​​control unit (750) forms a closed-loop structure in which control precision improves over time through active learning feedback. Fourth, the AI ​​control unit (750) functions as a cross-module of the integrated system to ensure the safety management output quality of the integrated system is above a certain level even with future AI technology evolution (NER → LLM → Agent → next-generation AI).

[0082] Example 4 ― Closed-loop reflux structure

[0083] FIG. 10 is a diagram illustrating data flow and closed-loop recirculation structures within an integrated system according to an embodiment of the present invention. Referring to FIG. 10, the integrated system according to an embodiment of the present invention forms three closed-loop recirculations.

[0084] The first feedback loop is the feedback from (e) the distributed learning unit (500) to the KG generation unit (300). The distributed learning unit (500) calculates the Bayesian-updated causal strength W_post from the workplace unit learning results and feeds it back to the spatiotemporal causal KG of the KG generation unit (300). As a result, the 1st-degree attribute W of the KG's causal strength is updated over time by reflecting the workplace data.

[0085] The second feedback loop is the feedback from (f) HVD curation unit (600) → (c) KG generation unit (300). The HVD curation unit (600) evaluates new external data (e.g., new chemical information, updated public datasets) using five quality indicators, registers them as HVDs, and feeds them back to the KG generation unit (300) as new nodes or new relationships. This expands the set of nodes in the KG over time.

[0086] The third feedback loop is the feedback from (g) the AI ​​control unit (750) to the AI ​​extraction and inference unit (200). The human reviewer labeling result of the AI ​​control unit (750) is fed back to (b) the AI ​​extraction and inference unit (200) as an active learning signal, and the extraction precision improves over time. Feedback is applied in the form of (b-1) model retraining in the case of the NER modality, (b-2) prompt few-shot updates in the case of the LLM modality, and (b-3) tool call policy updates in the case of the Agent AI modality.

[0087] The three recirculation loops mentioned above operate independently of each other while cumulatively contributing to the improvement of output precision of the integrated system. In particular, the third recirculation loop (AI control unit → AI extraction / inference unit) operates regardless of the type of AI modality, so the precision improvement mechanism of the integrated system is maintained regardless of future AI technology evolution.

[0088] Example 5 ― Application Scenario of 8 Major Components

[0089] In this embodiment, a scenario in which the integrated system of the present invention is applied to an actual chemical accident report is specifically exemplified.

[0090] [Input Accident Report] “Around 10:30 AM on March 15, 2024, hydrochloric acid leaked from a storage tank at the OO Chemical Plant in the Sihwa Industrial Complex, Ansan-si, Gyeonggi-do. Four nearby workers complained of difficulty breathing; the employer evacuated the workers to a safety zone starting at 10:45 AM, and the emergency response team carried out measures to stop the leak and neutralize the situation.”

[0091] [Step 1 ― Data Input Section (10)] The above accident report is received as natural language text.

[0092] [Step 2 ― AI Extraction / Inference Unit (200)] In this scenario, it is assumed that the (b-2) LLM modality is adopted. LLM outputs 5 entities as follows according to the extraction prompt: CHEM=“Hydrochloric acid”, FACI=“OO Chemical Plant Storage Tank”, ACCD=“Leak”, DAMG=“4 workers in respiratory distress”, RESP=“Evacuate·Block·Neutralize”. Additionally, LLM infers and outputs CAS=“7647-01-0”, GHS=“H314, H335”, and KOSHA=“CHM-7351”.

[0093] [Step 3 ― AI Control Unit (750)] The AI ​​Control Unit performs quadruple verification on the LLM output. (g-1) SHACL: Response-Effect relationship time order constraint (t_response 10:45 > t_accident 10:30) Passed. (g-2) Regulations: “Hydrochloric acid” corresponds to a hazardous chemical under the Chemical Substances Control Act and is mapped to the accident type “Leak” under the Dangerous Goods Safety Control Act ― Passed. (g-3) Triple Identifier: LLM output CAS 7647-01-0, GHS H314 / H335, KOSHA CHM-7351 matches an external standard dataset ― Passed. (g-4) Hallucination Detection: Identifier match 1.0, integrated uncertainty score u(x) = 0.08 (θ_AI = less than 0.40) ― Passed. Thus, the LLM output is processed as pass-through and transmitted to (c) KG generation unit (300).

[0094] [Step 4 ― KG generation unit (300)] A verified RDF triple is assigned spacetime first-degree attributes (τ, W, β, C) and registered as a new instance of spacetime causation KG.

[0095] [Step 5 ― Mixed Risk Calculation Unit (400)] SPARQL causal paths are extracted from the KG to calculate the composite risk score R. In this case, it is assumed that R = 1.85 is calculated.

[0096] [Step 6 ― Distributed Learning Unit (500) and HVD Curation Unit (600)] The case data is input into the Distributed Learning Unit (500) after obtaining consent from the workplace, and new accident patterns are fed back into the KG through the HVD Curation Unit (600).

[0097] [Step 7 ― Alarm Issuance Unit (700)] Since R = 1.85 is greater than or equal to the threshold value θ1 = 1.0 and less than θ2 = 2.5, it is classified as medium risk (YELLOW) and a caution alarm is issued. The integrated safety management output unit (800) transmits the alarm to the workplace safety manager.

[0098] [Application Effect ― AI Modality Independence] In this scenario, even if (b) the AI ​​extraction / inference unit (200) is adopted as (b-1) NER modality or (b-3) Agent AI modality, the AI ​​control unit (750) verifies the output using the same quadruple verification protocol, and subsequent steps of the integrated system (KG generation, risk calculation, and alarm issuance) operate identically. Thus, the present invention is AI modality neutral, and the quality of the safety management output of the integrated system is guaranteed to be above a certain level even with future AI technology evolution.

[0099] Comparative Example ― Comparison with the case where some or all of the components of the present invention are lacking

[0100] In order to objectively verify the effects of the present invention, Comparative Examples 1 to 4, which lack some or all of the components of the present invention, were designed and the safety management output quality was compared and measured on the same dataset (approximately 5,000 Korean chemical accident reports from 2020 to 2024).

[0101] Comparative Example 1 ― Absence of AI Control Unit (LLM Standalone Operation)

[0102] Comparative Example 1 is a case in which (g) the AI ​​control unit (750) of the present invention is removed, and (b) the LLM modality is operated alone as the AI ​​extraction and inference unit (200). (a) the integrated ontology unit, (c) the KG generation unit, (d) the mixed risk calculation unit, (e) the distributed learning unit, (f) the HVD curation unit, and (h) the alarm issuance unit are applied in the same way as in the present invention.

[0103] When Comparative Example 1 was applied to the same dataset, hallucinations (such as the generation of non-existent CAS numbers, GHS classifications inconsistent with chemical names, and response times contradicting the time of the accident) occurred in approximately 23% of the extraction results generated by the LLM; however, since this Comparative Example lacks an AI control unit, the hallucination results were directly incorporated into the KG and risk calculation. Consequently, the false alarm rate of the alarm issuing unit increased by approximately 4.2 times compared to the present invention. Comparative Example 1 demonstrates that the AI ​​control unit of the present invention is essential for ensuring the quality of safety management outputs in an integrated system.

[0104] Comparative Example 2 ― Modality limitation of AI extraction / inference unit (NER alone)

[0105] Comparative Example 2 is a case in which the (b) AI extraction / inference unit (200) of the present invention is limited to the NER modality alone, and the LLM modality and Agent AI modality are excluded. Other components including the AI ​​control unit (750) are applied in the same way as the present invention.

[0106] As a result of applying Comparative Example 2 to the same dataset, NER standalone operation showed extraction performance similar to the present invention for chemical accident reports with sufficient labeling data, but for reports of various formats or reports containing new chemical names, the extraction omission rate was about 35% higher than that of the present invention. Comparative Example 2 demonstrates that the generalization of modalities in the AI ​​extraction and inference unit contributes to expanding the scope of application of the integrated system.

[0107] Comparative Example 3 ― Recirculation Closed Loop Member

[0108] Comparative Example 3 is a case in which all three recirculation loops of the present invention (distributed learning → KG, HVD → AI control unit → AI extraction / inference unit) are removed, and the eight major components operate only as a unidirectional pipeline. As a result, the output quality was similar to that of the present invention in the initial learning rounds, but as the rounds progressed, the improvement in precision stagnated, and at the end of 100 rounds, the precision was approximately 18% lower than that of the present invention. Comparative Example 3 demonstrates that a closed-loop recirculation structure is an essential component for the improvement of precision over time.

[0109] Comparative Example 4 ― Inverse Application (Reversal of verification direction of AI control unit)

[0110] Comparative Example 4 is a case where the verification results of the AI ​​control unit of the present invention are applied in reverse, which is an inappropriate mapping in which results with an integrated uncertainty score u(x) above a threshold are automatically passed, and results below the threshold are escalated to human reviewers. As a result, 100% of the hallucinogenic output was input into the KG and risk calculation, and normal output was loaded into the reviewer queue, causing an excessive burden on human reviewers, and the false alarm rate of the alarm issuing unit increased by approximately eight times compared to the present invention. Comparative Example 4 demonstrates the essential meaning of the verification direction of the AI ​​control unit of the present invention.

[0111] Comprehensive comparative analysis of examples and comparative examples

[0112] The results of applying the embodiments of the present invention and comparative examples 1 to 4 to the same dataset and comparing the four major evaluation indicators (extraction accuracy, hallucination detection rate, false alarm rate, and AI modality independence) are shown in the following comparative analysis table.

[0113] Comprehensive comparison of examples and comparative examples division Extraction accuracy Hallucination detection rate False alarm rate AI modality irrelevance Examples (The present invention) About 92% About 96% About 4% guarantee Comparative Example 1 (Absence of AI control unit) About 88% 0% (Lack of verification) About 17% Uninsured Comparative Example 2 (NER alone) About 78% About 96% About 5% Not applicable (NER only) Comparative Example 3 (Refrigeration Absence) About 75% (stagnation) About 96% About 6% guarantee Comparative Example 4 (Reverse Application) About 92% 0% (Reverse) About 32% Non-guaranteed (misleep learning)

[0114] As shown in the comparative analysis table above, the embodiment of the present invention exhibits superior performance compared to Comparative Examples 1 to 4 in all four major evaluation indicators. In particular, regarding the hallucination detection rate, the embodiment shows approximately 96%, whereas Comparative Example 1 (absence of AI control unit) drops sharply to 0%. This demonstrates that the AI ​​control unit has a decisive influence on ensuring the quality of the safety management output of the integrated system. Furthermore, regarding AI modality independence, the present invention ensures independence by verifying the output with the same control protocol regardless of whether the NER modality, LLM modality, or Agent AI modality is adopted, whereas Comparative Example 2 (NER alone) has its modality fixed to NER and cannot be applied to LLM or Agent AI environments. Through the above comparative analysis, it was confirmed that the core differentiating components of the present invention (integration of 8 major components, generalization of AI extraction / inference unit modality, AI control unit, and closed-loop feedback) each contribute independently, and that maximum performance is achieved in the embodiment of the present invention which combines all of them.

[0115] Subsequent system linkage

[0116] The integrated system of the present invention can operate in combination with six components (A-2, B-1, B-2, B-3, B-4, B-5) according to the applicant's separate Korean patent application.

[0117] [Link 1] A-2 System (Integrated Ontology Schema): (a) The integrated ontology unit (100) of the present invention may utilize the ontology and SHACL integrity constraints defined by the A-2 system of the applicant’s separate Korean patent application. In particular, (g-1) SHACL verification unit (752) of (g) AI control unit (750) of the present invention may apply the SHACL Shape of the A-2 system as is. However, the present invention may be applied to any chemical safety ontology and is not limited thereto.

[0118] [Link 2] B-1 System (Space-time Causal KG): The (c) KG generation unit (300) of the present invention may utilize the space-time first-degree attribute assignment method and Bayesian causal strength update mechanism of the B-1 system, but the present invention is applicable to any space-time causal knowledge graph and is not limited thereto.

[0119] [Link 3] B-2 System (Korean Language Accident Report 5 Entity NER): The (b-1) NER modality (210) of the (b) AI extraction / inference unit (200) of the present invention may utilize the 5-entity simultaneous NER model, BIO tagging, CRF decoding, and domain adaptation backbone of the B-2 system. However, the (b) AI extraction / inference unit of the present invention is not limited to the NER modality and may also simultaneously adopt LLM or Agent AI modality.

[0120] [Link 4] B-3 System (KG Path-based Complex Risk Prediction): The (d) complex risk calculation unit (400) of the present invention may utilize the SPARQL causal path extraction and synergy score calculation formula of the B-3 system, but the present invention is applicable to any synergy score calculation means and is not limited thereto.

[0121] [Link 5] B-4 system (trade secret differential ε-distributed federated learning): The distributed learning unit (500) of the present invention (e) may utilize differential ε mapping, two-stage aggregation, and Bayesian W feedback of the B-4 system, but the present invention is applicable to any distributed learning system and is not limited thereto.

[0122] [Link 6] B-5 System (HVD Automatic Curation): The (f) HVD curation unit (600) of the present invention may utilize the calculation of five major quality indicators and weights for each policy area of ​​the B-5 system, but the present invention is applicable to any curation means and is not limited thereto.

[0123] [Integrated System Combination] The present invention protects the synergy of six separate applications as a single system claim combined with (g) an AI control unit (750), so that the present invention has independent rights value that does not overlap with the separate applications of individual modules. In particular, even if a new AI modality emerges due to future AI technology evolution, the (g) AI control unit (750) of the present invention operates as a modality-independent control protocol, so the scope of rights of the integrated system automatically adapts to the AI ​​technology evolution. Explanation of the symbols

[0125] 10: Data Input Section 100: Integrated Ontology Part ((a) Components) 200: AI Extraction / Inference Unit ((b) Components) 210: NER modality ((b-1)) 220: LLM modality ((b-2)) 230: Agent AI Modality ((b-3)) 240: Extraction and Inference Result Integration Unit 300: KG generation unit ((c) component unit) 400: Mixed Risk Calculation Unit ((d) Components) 500: Distributed learning unit ((e) component) 600: HVD Curation Department ((f) Composition Department) 700: Alarm issuance unit ((h) component) 750: AI Control Unit ((g) Components) 751: Input Integration Section 752: SHACL Integrity Verification Unit ((g-1)) 753: Legal Consistency Verification Department ((g-2)) 754: Triple Identifier Mapping Match Verification Unit ((g-3)) 755: Hallucination Detection / Uncertainty Score Calculation Unit ((g-4)) 756: Human Reviewer Escalation and Active Learning Feedback Unit ((g-5)) 757: Control Result Output Section 800: Integrated Safety Management Output Unit

Claims

Claim 1 As a knowledge graph-based chemical safety early warning integrated system, comprising: (a) an integrated ontology unit providing an integrated ontology that defines six object classes (Chemical, Accident, HealthImpact, Workplace, Worker, Environment) and four major relationship classes (causedBy, leadsToImpact, locatedIn, exposedTo) of the chemical safety domain; (b) an AI extraction and inference unit that extracts or infers five entities—chemical name, accident type, time, place, and symptom—from natural language text of a Korean accident report and converts them into RDF triples — comprising at least one of a pre-trained NER model, a large-scale language model (LLM), and an Agent AI system having a tool call function —; (c) a KG generation unit that generates a spatiotemporal causal knowledge graph by assigning spatiotemporal first-degree attributes (τ time lag, W causal strength, β spatial decay coefficient, C context vector) to the validated RDF triples; and (d) a composite risk score R = HI + by calculating causal path weighted synergy from a chemical set and the KG. (e) a mixed risk calculation unit that calculates α × S_total; (e) a distributed learning unit that extracts workplace adaptation parameters using distributed federated learning with differential privacy ε based on workplace trade secret (CBI) grades and feeds back Bayesian-updated causal strengths to the KG generation unit; (f) an HVD curation unit that evaluates external data candidates using five quality indicators (completeness, accuracy, timeliness, consistency, and accessibility) to automatically curate high-value data (HVD) and feeds back to the KG generation unit;(g) an AI control unit that performs SHACL integrity constraint verification, legal compliance verification according to the statutory classification of the Chemical Substances Control Act, the Occupational Safety and Health Act, and the Dangerous Goods Safety Management Act, CAS, GHS, and KOSHA triple identifier mapping consistency verification, and calculates an integrated uncertainty score by combining hallucination detection signals on the output result of the AI ​​extraction and inference unit of (b) above, and automatically escalates results where the integrated uncertainty score is above a predefined threshold to a human reviewer queue and feeds back the reviewer labeling results as active learning signals to the AI ​​extraction and inference unit of (b) above; and (h) an alarm issuance unit that issues a three-stage alarm (green / yellow / red) by synthesizing the composite risk score R and the accident scenario DB matching result; comprising a knowledge graph-based chemical safety early warning integrated system.; Claim 2 The knowledge graph-based chemical safety early warning integrated system according to claim 1, wherein (a) the integrated ontology part may utilize an ontology defined by the chemical safety integrated ontology schema of the applicant’s separate Korean patent application (hereinafter referred to as the 'A-2 system'), but the present invention is applicable to any chemical safety ontology and is not limited thereto. Claim 3 The knowledge graph-based chemical safety early warning integrated system according to claim 1, wherein the (b) AI extraction and inference unit is implemented with a pre-trained Korean NER model, and may apply 5-entity simultaneous NER, BIO tagging, and CRF decoding by the Korean accident report named entity recognition system of the applicant's separate Korean patent application (hereinafter referred to as 'B-2 system'), but the present invention is applicable to any Korean NER model and is not limited thereto. Claim 4 A knowledge graph-based chemical safety early warning integrated system according to claim 1, wherein the (b) AI extraction and inference unit is implemented with a pre-trained large-scale language model (LLM), and outputs 5 entities from a Korean accident report in a structured output format by means of an extraction prompt including class definitions of an integrated ontology and name and format specifications of 5 entities. Claim 5 The knowledge graph-based chemical safety early warning integrated system according to claim 1, wherein the (b) AI extraction and inference unit is implemented as an Agent AI system having a tool calling function, and performs multi-stage inference by calling at least one external tool among a Korean morphological analyzer, a CAS·GHS·KOSHA identifier lookup unit, a chemical name normalizer, and a hazardous material type classifier. Claim 6 The knowledge graph-based chemical safety early warning integrated system, wherein the (c) KG generation unit of the present applicant may utilize the spatiotemporal causal knowledge graph system of a separate Korean patent application (hereinafter referred to as the 'B-1 system'), but the present invention is applicable to any spatiotemporal causal knowledge graph and is not limited thereto. Claim 7 The knowledge graph-based chemical safety early warning integrated system, wherein the (d) mixing risk calculation unit extracts a set of causal pathways connecting chemical pairs using a SPARQL query and calculates a synergy score S(Ci, Cj) = Σw(Pk) × Π∈g(τ, W, β, C), and may utilize the knowledge graph path-based complex chemical mixture exposure risk prediction system (hereinafter referred to as 'B-3 system') of the applicant's separate Korean patent application, but the present invention is applicable to any synergy score calculation means and is not limited thereto. Claim 8 In claim 1, the distributed learning unit (e) classifies workplace trade secret (CBI) grades into three levels: Cat-A, Cat-B, and Cat-C, maps differential privacy budgets to each grade in a ratio of ε_A: ε_B: ε_C = 1:2:4, and updates a global model through a two-stage aggregation of a weighted average by grade and a weighted sum between grades, wherein the workplace-unit distributed federated learning system (hereinafter referred to as the 'B-4 system') of the applicant's separate Korean patent application may be utilized, but the present invention is applicable to any distributed learning system and is not limited thereto, a knowledge graph-based chemical safety early warning integrated system. Claim 9 In claim 1, the (f) HVD curation unit calculates the average of the five quality indicators Q = (Q_comp + Q_acc + Q_time + Q_cons + Q_acce) / 5 and automatically registers data that is greater than or equal to a threshold value determined in the range of 0.7 to 0.9 as HVD, and may utilize the chemical safety HVD automatic curation system of the applicant's separate Korean patent application (hereinafter referred to as 'B-5 system'), but the present invention is applicable to any curation means and is not limited thereto, a knowledge graph-based chemical safety early warning integrated system. Claim 10 In claim 1, the (g) AI control unit comprises: (g-1) a SHACL integrity verification unit that verifies whether there is a violation of cardinality, datatype, value range, referential integrity, chronological order, and facility hierarchy constraints by applying the SHACL Shape of the integrated ontology to the output 5 entities of the (b) AI extraction and inference unit; (g-2) a regulatory consistency verification unit that verifies consistency according to the statutory classification system of the Chemical Substances Control Act, the Occupational Safety and Health Act, and the Dangerous Goods Safety Management Act; (g-3) a triple identifier mapping consistency verification unit that verifies whether the triple identifiers of the CAS number, GHS classification code, and KOSHA code for the extracted chemical substance match an external standard dataset; and (g-4) a weighted sum of the verification results of (g-1) to (g-3) to calculate an integrated uncertainty score u(x) = α·u_SHACL(x) + β·u_regulation(x) + γ·u_identifier(x) + δ·u_modality(x). A knowledge graph-based chemical safety early warning integrated system comprising: a hallucination detection and uncertainty score calculation unit ― where α and βγδ are normalized as weights so that their sum is 1 ―; and (g-5) a human reviewer escalation and active learning feedback unit that automatically loads extraction results in which the integrated uncertainty score is greater than or equal to a predefined threshold θ into a human reviewer queue and feeds the reviewer labeling results back as active learning signals to the (b) AI extraction and inference unit. Claim 11 A knowledge graph-based chemical safety early warning integrated system, wherein, in claim 10, the threshold value θ_AI is determined in the range of 0.30 or higher and 0.50 or lower, such that if θ_AI is less than 0.30, almost all extraction results are escalated to human reviewers, thereby increasing the reviewer's burden, and if θ exceeds 0.50, hallucinogenic results are automatically passed through and errors are introduced into the safety management output, so the range of 0.30 or higher and 0.50 or lower has critical significance corresponding to the balance point between the reviewer's burden and the output quality. Claim 12 A knowledge graph-based chemical safety early warning integrated system according to claim 10, wherein the (g) AI control unit verifies, filters, and feeds back the output using the same control protocol regardless of whether the (b) AI extraction and inference unit is implemented as a NER model, LLM, or Agent AI, thereby ensuring the safety management output quality of the integrated system at a level above a certain standard regardless of the type of AI modality and future AI technology evolution. Claim 13 The knowledge graph-based chemical safety early warning integrated system according to claim 1, wherein the (h) alarm issuing unit compares the composite risk score R with a threshold value determined within the range of a first threshold value θ = 0.7 or higher and 1.3 or lower and a second threshold value θ = 1.8 or higher and 3.5 or lower, classifies it into three levels as low risk (GREEN) if R is less than θ, medium risk (YELLOW) if R is θ or higher and less than θ, and issues an alarm corresponding to each level. Claim 14 A knowledge graph-based chemical safety early warning integrated system, wherein the three feedback loops of (e) Bayesian updated causal strength feedback of the distributed learning unit, (f) new data listing feedback of the HVD curation unit, and (g) human reviewer labeling result feedback of the AI ​​control unit are each applied to (c) KG generation unit and (b) AI extraction and inference unit to form a closed-loop feedback structure, and the extraction precision and safety management output quality of the integrated system increase monotonically over time. Claim 15 The knowledge graph-based chemical safety early warning integrated system according to claim 1, wherein the integrated system is capable of single-site operation or multiple-site consortium operation and is linked with external systems through SPARQL API and REST API. Claim 16 A knowledge graph-based chemical safety early warning integrated method comprising: (1) a step in which a data input unit receives accident reports, measurement data, chemical specifications, and environmental condition data; (2) a step in which an AI extraction and inference unit extracts or infers five entities—chemical name, accident type, time, location, and symptoms—from the input Korean accident report and converts them into RDF triples, wherein this is performed by at least one of a pre-trained NER model, a large-scale language model (LLM), and an Agent AI system; (3) a step in which an AI control unit performs SHACL integrity verification, regulatory compliance verification, and triple identifier mapping agreement verification on the output result of step (2) to calculate an integrated uncertainty score, escalates results above a threshold to a human reviewer queue, and feeds back the reviewer labeling result to step (2) as an active learning signal; (4) a step in which a KG generation unit generates a spatiotemporal causal knowledge graph by assigning spatiotemporal first-degree attributes τW and β to the RDF triples that have passed control; (5) a mixed risk calculation unit [calculates] causal path weighted synergy using a SPARQL query A knowledge graph-based chemical safety early warning integrated method comprising: a step of calculating a composite risk score R = HI + α × S_total; (6) a step in which a distributed learning unit performs differential ε-variance federated learning according to the trade secret grade of the workplace to extract workplace adaptation parameters and feed back Bayesian updated causal strength to the KG, and an HVD curation unit evaluates external data using five major quality indicators and registers it as a new entry in the KG; and (7) a step in which an alarm issuing unit combines the results of matching R with accident scenarios to issue a three-stage alarm. Claim 17 A knowledge graph-based chemical safety early warning integrated method according to claim 16, wherein steps (4) to (6) are combined into a closed-loop reflux structure, the distributed learning result and HVD curation result of step (6) are refluxed to the KG of step (4), and the review result of the AI ​​control unit of step (3) is refluxed to the AI ​​extraction and inference unit of step (2) so that an improvement in accuracy over time is achieved. Claim 18 A computer-readable recording medium having a program recorded thereon for executing the method of claim 16 or claim 17 on a computer.

Images