Integrated safety management method, apparatus, and system for ai-based prior identification of work hazard and calculation of dynamic safety score

Abstract

One embodiment of the present invention relates to an AI-based integrated safety management method, apparatus, and system for pre-identifying work risks and calculating dynamic safety scores, which predicts risks in advance by fusion-analyzing heterogeneous multimodal data from an industrial site and derives a quantitative integrated safety score based thereon.

Description

Technology Field

[0001] The following embodiments relate to an AI-based integrated safety management method, apparatus, and system for pre-identifying work risks and calculating dynamic safety scores, which predicts risks in advance by fused and analyzing heterogeneous multimodal data from industrial sites and derives quantitative integrated safety scores based thereon. Background Technology

[0002] Recently, various safety equipment and systems, such as CCTVs and IoT sensors, have been introduced at industrial sites to prevent serious accidents. However, as individual systems operate independently and in a fragmented manner, there have been limitations in early detection of hazardous situations caused by complex factors and in integrated monitoring.

[0003] There is a lack of technical means to predict the likelihood of accidents by comprehensively analyzing the interaction of unsafe worker behaviors, on-site environmental factors, equipment conditions, and managerial elements (work permits, safety inspection documents, meeting contents, etc.).

[0004] Furthermore, due to the absence of quantitative indicators that can reflect the actual safety level at the site in real time and evaluate it using intuitive and objective figures, managers are struggling to immediately identify changing risks at the site and take preemptive control measures. Prior art literature

[0005] (Patent Document 0001) KR 10-2910717 B (Patent Document 0002) KR 10-2284691 B (Patent Document 0003) KR 10-1723283 B The problem to be solved

[0006] The problem that an embodiment of the present invention aims to solve is to provide an AI-based integrated safety management method, apparatus, and system for pre-identifying work risks and calculating dynamic safety scores, which overcomes the limitations of the conventional fragmented safety monitoring system and the absence of quantitative evaluation indicators as described above by integrating and analyzing multimodal data generated at the site using AI technology to preemptively identify the possibility of accidents and deriving real-time dynamic safety scores by combining multidimensional indicators from the perspectives of workers, the site, and the organization. means of solving the problem

[0007] According to one embodiment, an integrated safety management method for AI-based pre-identification of work risks and calculation of dynamic safety scores, wherein a device includes a processor, memory, a communication module, and a non-transient storage medium, and the processor executes a program stored in the non-transient storage medium, comprises: a step of acquiring a plurality of work-related data including IoT environment sensor data, CCTV video data, work plan data, work permit data, meeting audio data, safety inspection documents, and equipment log data from a work site; a step of generating integrated input data for analysis by performing preprocessing on the work-related data including data format unification, noise removal, missing value correction, time synchronization, and normalization; a step of deriving work risk prediction information including risk factors by work process, worker unsafe behavior, abnormal conditions of the site environment, signs of equipment abnormality, and the possibility of accident occurrence by performing multimodal fusion analysis on the integrated input data for analysis using a pre-trained artificial intelligence model; and a step of calculating worker behavior indicators, site condition indicators, and organizational management indicators based on the work risk prediction information. The present invention provides an AI-based integrated safety management method for pre-identifying work risks and calculating dynamic safety scores, comprising the step of calculating an integrated safety score by applying pre-set weights to the worker behavior indicators, site condition indicators, and organizational management indicators, and assigning a safety grade corresponding to the integrated safety score.

[0008] In addition, the step of deriving the above-mentioned work risk prediction information comprises: a step of deriving abnormal field environmental conditions and signs of equipment abnormalities, and extracting time-series latent vectors by applying a prediction model based on a Long Short-Term Memory (LSTM) algorithm to time-series analysis of patterns such as surges in hazardous gas concentration, exceedance of temperature and humidity thresholds, and abnormal equipment vibration within a preset time window for time-series stream data including IoT environment sensor data and equipment log data among the above-mentioned integrated input data for analysis; and a step of identifying unsafe worker behaviors and extracting spatial behavioral latent vectors by applying a vision AI model, which combines object detection and 3D skeleton estimation algorithms, to CCTV video data among the above-mentioned integrated input data for analysis to analyze spatiotemporal behavioral patterns including the worker's non-wearing of personal protective equipment (PPE), unsafe movement paths, entry into hazardous areas, and failure to maintain a safe distance from heavy equipment. A step of extracting risk factors for each work process, including inherent potential risks of the work process on the day, and extracting text embedding vectors by performing named entity recognition (NER) and contextual analysis using an industrial safety-specific pre-trained language model (Domain-Specific Transformer) on text data converted from work plan data, work permit data, and meeting voice data, and safety inspection documents among the integrated input data for analysis; a step of simultaneously inputting the time-series potential vector, spatial behavioral potential vector, and text embedding vector into a Cross-Attention network layer to generate a single multi-modal fused vector reflecting the mutual correlation weights between heterogeneous data;The method may include the step of inputting the above multimodal fusion vector into a gradient boosting-based ensemble classifier (XGBoost) to calculate the probability of accident occurrence by type of major accident, including falls, entrapment, collision, electric shock, fire, explosion, and suffocation, at a specific point in time at a target work site as a probability index in percentage units; and the step of generating work risk prediction information by combining the said probability of accident occurrence with risk factors, unsafe behaviors, abnormal conditions of the site environment, and signs of equipment abnormality.

[0009] And, the step of calculating worker behavior indicators, site condition indicators, and organizational management indicators based on the above work risk prediction information comprises: generating risk event records by mapping worker identification information, work process identification information, work zone identification information, equipment identification information, and accident type identification information for each of the work process-specific risk factors, unsafe worker behaviors, abnormal site environment conditions, signs of equipment abnormalities, and accident probability included in the above work risk prediction information; clustering the risk event records based on worker units, work process units, work zone units, equipment units, and time window units, respectively, to generate risk event sets by worker, risk event sets by work zone, risk event sets by work process, and risk event sets by equipment, respectively; and for each of the risk event sets by worker, calculating the frequency of non-wearing of protective equipment, frequency of unsafe movement paths, number of entry into hazardous zones, number of failure to maintain safety distances, duration of abnormal postures, average interval between repeated unsafe behaviors, simultaneous occurrence rate of multiple unsafe behaviors by the same worker, and weighted average values ​​of accident probability. A step of performing minimum-maximum normalization on the weighted average values ​​of the frequency of non-wearing of protective equipment, frequency of unsafe movement paths, number of entry into hazardous zones, number of failure to secure safety distances, duration of abnormal postures, average interval between repetitive unsafe behaviors, simultaneous occurrence rate of multiple unsafe behaviors by the same worker, and probability of accident occurrence, and calculating worker behavior indicators by applying risk type weights, work process weights, accident severity weights, and worker role weights to the normalized values;For each of the above-mentioned sets of risk events by work zone, sets of risk events by work process, and sets of risk events by equipment, a step of calculating the degree of rapid increase in hazardous gas concentration, the degree of exceeding temperature and humidity thresholds, the duration of abnormal conditions, the intensity of abnormal equipment vibration, the simultaneous occurrence rate of multiple abnormal conditions, the spread area of ​​the risk zone, the recurrence rate of abnormal equipment signs, and the degree of overlap of multiple high-risk events in the same time period; a step of performing minimum-maximum normalization on the degree of rapid increase in hazardous gas concentration, the degree of exceeding temperature and humidity thresholds, the duration of abnormal conditions, the intensity of abnormal equipment vibration, the simultaneous occurrence rate of multiple abnormal conditions, the spread area of ​​the risk zone, the recurrence rate of abnormal equipment signs, and the degree of overlap of multiple high-risk events in the same time period, and applying weights for environmental items, weights for work processes, weights for equipment importance, and weights for disaster severity to the normalized values ​​to calculate the environmental deterioration score by work zone, the condition instability score by work process, and the abnormal concentration score by equipment; and a step of calculating the site condition index by weighting the environmental deterioration score by work zone, the condition instability score by work process, and the abnormal concentration score by equipment. A step of calculating, for all of the above risk event records, the frequency of recurrence of the same type of accident, the ratio of simultaneous exposure of multiple workers, the ratio of simultaneous high risk of multiple work processes, the spread of risk between multiple work zones, the simultaneous occurrence rate of equipment anomalies and unsafe behaviors, the number of tasks requiring management intervention, and the concentration of high-risk events;The method may include the step of performing minimum-maximum normalization on the frequency of recurrence of the same type of accident, the ratio of simultaneous exposure of multiple workers, the ratio of simultaneous high risk of multiple work processes, the degree of risk diffusion between multiple work zones, the rate of simultaneous occurrence of signs of equipment abnormality and unsafe behaviors, the number of tasks requiring management intervention, and the concentration of high-risk events, and applying weights by accident type, work process importance, work zone impact, and the degree of need for management intervention to the normalized values ​​to calculate an organizational-level risk repeatability score, an organizational-level risk diffusion score, an organizational-level simultaneous management burden score, and an organizational-level complex risk management difficulty score; and the step of calculating an organizational management indicator by weighted summing the organizational-level risk repeatability score, the organizational-level risk diffusion score, the organizational-level simultaneous management burden score, and the organizational-level complex risk management difficulty score.

[0010] In addition, the method further comprises the steps of: acquiring voice data of a safety inspection meeting; performing voice preprocessing on the voice data, including noise removal, sound quality correction, and multi-speaker separation, and then generating meeting text using a speech recognition model; performing natural language processing-based keyword extraction and context analysis on the meeting text to derive meeting analysis information including work content, work process, equipment used, risk factors, safety instructions, and participating worker information; mapping the meeting analysis information with at least one of a pre-configured standard safety work procedure database, a risk assessment database, and a past accident case database to derive risk items and recommended safety measures corresponding to the work content; and automatically inputting content into a pre-configured TBM document template based on the risk items, recommended safety measures, and participating worker information to generate a pre-work safety inspection document; wherein the step of deriving the meeting analysis information includes: a step of first refining non-standard field terms, abbreviations, and colloquial expressions within the meeting text into standard safety terms by applying an industrial site-specific user dictionary, an expanded field abbreviation dictionary, and slang / dialect mapping metadata; A step of generating triplet data in which words or phrases within the text are related and structured into a triplet structure including an object (equipment used, tool, material, or work object), an action (work content, work sequence, or work process), and a condition (site environment, risk factor, or work condition) by dividing the primary refined meeting text into sentence and clause units, and then performing Named Entity Recognition, Dependency Parsing, and Semantic Role Analysis;A step of generating work-behavior-risk association metadata by additionally tagging the work location, work time zone, type of equipment used, worker role, hazard expression, safety instruction expression, and protective equipment-related expression for the above triplet data; a step of generating multidimensional embedding vectors by inputting the above work-behavior-risk association metadata into a pre-trained industrial safety specialized language model; a step of calculating the cosine similarity between the above multidimensional embedding vectors and the similar accident case sentence vectors, risk situation description vectors, and safety measure description vectors included in a pre-set industrial safety standard corpus, respectively, to select risk association sentence groups whose similarity scores are above a pre-set threshold; a step of generating a group of potential risk factors corresponding to at least one of falling, entrapment, collision, electric shock, overturning, falling objects, fire, explosion, suffocation, collapse, and hazardous substance exposure by analyzing the combination patterns of objects, behaviors, and states included in the risk association sentence groups; and a step of obtaining attendance verification data and work assignment information from a database. A step of extracting the name, affiliation, job title, and assigned task of a meeting participant using the meeting text, the speaker separation result derived from the voice preprocessing process, attendance confirmation data, and work assignment information, and generating participating worker information by obtaining qualifications, proficiency, recent work history, and safety training completion history from a database based on the name, affiliation, job title, and assigned task; a step of identifying imperative expressions, prohibitive expressions, confirmation request expressions, and attention-raising expressions from the work-behavior-risk association metadata, and classifying the identified imperative expressions, prohibitive expressions, confirmation request expressions, and attention-raising expressions into at least one safety instruction among instructions to wear protective equipment, instructions to prohibit access, instructions to comply with work sequence, instructions to inspect equipment, and instructions to check the environment; a step of obtaining external site environment information at the scheduled time of work and past accident history information of the target work process from a database.For each of the above-mentioned potential risk factor candidate groups, the method may include: a step of calculating a TBM reflection priority value for each of the above-mentioned potential risk factor candidate groups using a risk priority calculation model that applies individual worker proficiency, qualifications held, recent work history, safety training completion history, external site environment information at the scheduled time of work, and past accident history information of the target work process included in the above-mentioned participating worker information as weighting variables; a step of determining the item-by-item reflection order of caution phrases, items requiring verification, and preventive action phrases corresponding to the potential risk factor candidate groups based on the above-mentioned TBM reflection priority value; and a step of generating meeting analysis information including work content, work process, equipment used, risk factors, safety instructions, and participating worker information using the above-mentioned triplet data, potential risk factor candidate groups, safety instructions, and participating worker information.

[0011] In addition, the step of calculating the priority value for TBM reflection for each of the aforementioned candidate groups of potential risk factors comprises: a step of obtaining pre-calculated work risk prediction information from a database; a step of extracting work process-specific risk factors, worker unsafe behaviors, abnormal site environmental conditions, signs of equipment abnormalities, and the probability of accident occurrence included in the work risk prediction information; a step of mapping the work risk prediction information and the candidate groups of potential risk factors based on commonalities among work processes, work locations, equipment used, work time zones, and participating workers; a step of calculating a risk prediction correction value that reflects the probability of accident occurrence, whether unsafe behaviors apply, whether abnormal site environmental conditions apply, and whether signs of equipment abnormalities apply, based on the mapping results for each of the candidate groups of potential risk factors; and a step of calculating a basic priority value for each of the candidate groups of potential risk factors based on the individual worker's proficiency, qualifications held, recent work history, safety training completion history, external site environment information at the scheduled time of work, and past accident history information of the target work process included in the participating worker information. and may include the step of determining TBM reflection priority values ​​for each potential risk factor candidate group by applying the above risk prediction correction value to the priority base value with weights.

[0012] A device according to one embodiment may be combined with hardware and controlled by a computer program stored on a medium to execute the method of any one of the methods described above. Effects of the invention

[0013] According to one embodiment, by collecting multiple work-related data (IoT, video, unstructured documents, voice, etc.), performing preprocessing such as data format standardization, and then applying multimodal fusion analysis, complex and potential on-site risk factors that are difficult to identify from a single data source can be identified in advance with high accuracy.

[0014] Furthermore, by calculating indicators for workers, sites, and organizations based on this predictive information and quantifying them into integrated safety scores and grades, it enables managers to intuitively perceive the safety level of the site and perform systematic safety management.

[0015] Furthermore, by applying AI models optimized for time-series data (LSTM), image data (vision AI), and unstructured text data (specialized language models) respectively to extract individual latent vectors suited to data characteristics, and by fusing them into a Cross-Attention layer to analyze causal correlations, it is possible to predict the probability of accident occurrence by type of major accident with high precision, reflecting the organic interaction between heterogeneous data.

[0016] Furthermore, by finely clustering identified predictive information into units such as workers, work zones, processes, and equipment, and then performing min-max normalization on the detailed characteristics of each element (occurrence frequency, duration, overlap, etc.) and applying customized weights, it is possible to calculate individual worker behavioral indicators, field indicators representing the condition of physical spaces and equipment, and organizational indicators reflecting enterprise-wide management capabilities in a highly objective and dynamic manner. This enables the precise targeting of field vulnerabilities at the detailed level and allows for the establishment of effective disaster prevention measures.

[0017] Furthermore, by generating text through noise removal and multi-speaker separation of audio from safety inspection meetings, and immediately mapping this to standard safety work procedures and a database of past accident cases, it is possible to 혁신적으로 reduce the administrative burden of manual documentation for field managers and produce systematic documents that ensure no key safety measures are omitted.

[0018] Furthermore, the accuracy of natural language processing can be maximized by refining slang or dialects unique to industrial sites into standard safety terms and precisely analyzing the text in a triplet structure. In addition, by applying individual worker proficiency, the weather conditions of the day, and the past accident history of the target process as multivariate weighting variables to dynamically calculate the priority of reflection for each risk factor, customized key safety instructions that best suit the day's site conditions and worker condition can be placed at the top of the document, rather than in a formal document.

[0019] Furthermore, by cross-mapping objective work risk prediction information—such as abnormal environmental conditions, signs of equipment malfunction, and unsafe worker behaviors at the work site, which has been fused and analyzed in advance—with the meeting analysis results, converting this into risk prediction correction values, and finally reflecting them in the priority, it is possible to establish an advanced integrated safety management system that proactively incorporates real-time potential physical risks at the site into documents, rather than relying solely on the manager's verbal statements. Brief explanation of the drawing

[0020] FIG. 1 is a schematic diagram showing an AI-based work risk pre-identification and dynamic safety score calculation integrated safety management system according to an embodiment of the present invention. FIG. 2 is a flowchart illustrating an integrated safety management method for AI-based pre-identification of work risks and dynamic safety score calculation according to an embodiment of the present invention. FIG. 3 is a flowchart illustrating the steps for deriving work risk prediction information of an AI-based work risk pre-identification and dynamic safety score calculation integrated safety management method according to an embodiment of the present invention. FIG. 4 is a flowchart illustrating the steps for calculating worker behavior indicators, site condition indicators, and organizational management indicators of an AI-based work risk pre-identification and dynamic safety score calculation integrated safety management method according to an embodiment of the present invention. Specific details for implementing the invention

[0021] Hereinafter, embodiments are described in detail with reference to the attached drawings. However, various modifications may be made to the embodiments, and thus the scope of the patent application is not limited or restricted by these embodiments. It should be understood that all modifications, equivalents, and substitutions to the embodiments are included within the scope of the rights.

[0022] Specific structural or functional descriptions of the embodiments are disclosed for illustrative purposes only and may be modified and implemented in various forms. Accordingly, the embodiments are not limited to the specific disclosed forms, and the scope of this specification includes modifications, equivalents, or substitutions that fall within the technical concept.

[0023] Terms such as "first" or "second" may be used to describe various components, but these terms should be interpreted solely for the purpose of distinguishing one component from another. For example, the first component may be named the second component, and similarly, the second component may be named the first component.

[0024] When it is stated that a component is "connected" to another component, it should be understood that it may be directly connected to or coupled with that other component, or that there may be other components in between.

[0025] The terms used in the embodiments are for illustrative purposes only and should not be interpreted as intended to be limiting. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as "comprising" or "having" are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0026] In particular, where a 'step' in this specification is described as 'comprising' one or more detailed steps or sub-steps, said 'step' may be interpreted as including its own basic processing step while simultaneously performing the described detailed steps as well.

[0027] For example, if it is stated that 'a step of doing B to A' includes 'a step of doing D to C; a step of doing F to E; and a step of doing H to G,' the 'step of doing B to A' may be interpreted not merely as the basic operation of doing B to A, but as a configuration that performs detailed procedures together, such as a step of doing D to C, a step of doing F to E, and a step of doing H to G.

[0028] Accordingly, the above configuration does not exclude various sub-procedures included within the scope of execution of the corresponding step, and may be included within the scope of the present invention even if other procedures or means performing substantially the same or equivalent functions are substituted.

[0029] Expressions such as 'end part', 'both ends', 'one end', 'other end', and 'side end' of a component can be interpreted as referring to at least / any one of the end parts of that component.

[0030] In the description of the present invention, 'a method in which a device comprises a processor, a memory, a communication module, and a non-transient storage medium, and a program stored in the non-transient storage medium is executed by the processor,' the term 'method' may be interpreted as referring to the program stored in the non-transient storage medium itself or a part of the program.

[0031] The term 'Return' as used in the description of the present invention may refer to a result value being output, returned, or returned from a method, procedure, function, etc. used in a given program language / structure.

[0032] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which the embodiments pertain. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this application. For example, the term 'artificial intelligence model' may be selected from one or more of known general artificial intelligence models.

[0033] In addition, when describing with reference to the attached drawings, identical components are assigned the same reference numeral regardless of drawing symbols, and redundant descriptions thereof are omitted. In describing the embodiments, if it is determined that a detailed description of related prior art could unnecessarily obscure the essence of the embodiments, such detailed description is omitted.

[0034] According to one embodiment, a device includes a processor, memory, a communication module, and a non-transient storage medium, and a method for automatically generating a pre-work safety inspection (TBM) document using voice recognition-based safety meeting analysis, wherein the method is performed by the processor by executing a program stored in the non-transient storage medium, the method comprises: a step of acquiring voice data of a safety inspection meeting; a step of generating meeting text using a voice recognition model after performing voice preprocessing on the voice data, including noise removal, voice quality correction, and multi-speaker separation; a step of deriving meeting analysis information including work content, work process, equipment used, risk factors, safety instructions, and participating worker information by performing natural language processing-based keyword extraction and context analysis on the meeting text; and a step of mapping the meeting analysis information to at least one of a pre-established standard safety work procedure database, a risk assessment database, and a past accident case database to derive risk items and recommended safety measures corresponding to the work content. The present invention provides a method for automatically generating a pre-work safety inspection (TBM) document using voice recognition-based safety meeting analysis, comprising the step of generating a pre-work safety inspection document by automatically inputting content into a pre-configured TBM document template based on the above-mentioned risk items, recommended safety measures items, and participating worker information.

[0035] The step of acquiring voice data from safety inspection meetings is a step in which a device collects voice signals spoken by workers and site managers during TBM meetings conducted before the start of work and secures them in the form of digital data capable of subsequent analysis.

[0036] Here, the term "device" may be implemented as a field server, cloud server, edge computing device, administrator terminal, or a combination thereof, and may include a processor, memory, a communication module, and a non-transient storage medium. The processor may refer to a computational processing unit that performs operations related to the collection, analysis, and document generation of conference audio; the memory may refer to a storage area that stores voice data, intermediate analysis results, and generated documents; and the communication module may be a component that performs communication with a microphone device, a conference room terminal, a worker terminal, an administrator terminal, or an external database. Additionally, the non-transient storage medium may refer to a storage medium that stores program instructions, pre-trained models, database access information, and document templates.

[0037] A safety inspection meeting may refer to a pre-work safety meeting, or toolbox meeting, conducted to share and familiarize oneself with work details, risk factors, equipment to be used, protective gear, and precautions before commencing work on the day.

[0038] The above voice data may be acquired in real time through a voice collection means equipped in a fixed microphone installed in a conference room, a wearable microphone worn by a worker or safety manager, a mobile conference recording terminal, or a field manager terminal, or may be collected by uploading it to a server in the form of an audio file after recording is completed. For example, the device may automatically start recording in response to the start time of the meeting or start recording in response to a meeting start command entered from the manager terminal. At this time, the voice data may be collected in the form of waveform data of a mono channel or stereo channel, a compressed audio file, or a real-time streaming packet, and the meeting date and time, work site identification information, work team information, meeting organizer information, and collection device identification information may be mapped together with each voice segment.

[0039] The step of generating meeting text using a speech recognition model after performing speech preprocessing on the above speech data, including noise removal, sound quality correction, and multi-speaker separation, is a step of improving the quality of the collected speech data, distinguishing the utterances of multiple speakers, and then converting them into text data.

[0040] Here, noise removal may refer to a process that eliminates or attenuates acoustic components that interfere with the interpretation of meeting content, such as construction equipment sounds, wind noise, vehicle movement sounds, metal striking sounds, background conversation sounds, and communication noise. Additionally, sound quality correction may refer to a process that improves speech recognition accuracy by performing volume deviation correction, voice band emphasis, distortion correction, echo removal, or voice segment amplification. Furthermore, multi-speaker separation may refer to a process that analyzes the voice characteristics of speakers within mixed voice data to distinguish utterances by speaker—such as those of managers, workers, and signalmen—and associates each utterance with its respective speaker.

[0041] The step of generating conference text using a speech recognition model after performing speech preprocessing including noise removal, sound quality correction, and multiple speaker separation on the above speech data can be performed, for example, by first detecting silent and speech segments of the speech data, applying a noise removal filter for each speech segment, performing automatic gain control for segments with low volume, then separating sequences by speaker using a speaker separation model or a speaker segmentation algorithm, and then inputting each sequence into a speech recognition model to convert it into text.

[0042] At this time, the above-mentioned speech recognition model may include a known automatic speech recognition model, namely an ESTITY model, and may be a field-specific speech recognition model that has additionally learned equipment names, process names, hazard expressions, and protective equipment terms frequently used in industrial safety sites. For example, if a speech such as “We are starting ceiling tray work in Zone 3 today, and please check your safety belts before raising the aerial work platform” is collected during a meeting, the device can convert it into text and, if necessary, save it in the form of a text script recorded in chronological order along with a speaker identification value.

[0043] Here, meeting text may refer to character-based meeting record data generated by the above-mentioned speech recognition model.

[0044] The step of deriving meeting analysis information, including work content, work process, equipment used, risk factors, safety instructions, and participating worker information, by performing natural language processing-based keyword extraction and contextual analysis on the above meeting text, is a step of extracting key safety management elements related to the day's work from the generated text script and generating them as structured data.

[0045] Here, "work content" may refer to the purpose or target task of the specific work to be performed on the day; "work process" may refer to the procedural classification or location within the process; and "equipment used" may refer to the equipment, tools, machinery, or means of transport used to perform the work. Additionally, "risk factors" may refer to elements that cause accidents or increase the likelihood of safety accidents during the work process; "safety instructions" may refer to instructions, warnings, prohibitions, or requests for verification that must be checked or observed before work begins; and "participating worker information" may refer to personnel composition information, such as workers, managers, signalmen, and equipment operators, attending meetings or assigned to the work on the day.

[0046] The step of deriving meeting analysis information including work content, work process, equipment used, risk factors, safety instructions, and participating worker information by performing natural language processing-based keyword extraction and context analysis on the above meeting text can be implemented by dividing the meeting text into sentence units or utterance units, and then performing morphological analysis, part-of-speech tagging, named entity recognition, dependency parsing, semantic role analysis, keyword importance calculation, and context classification.

[0047] Natural language processing-based contextual analysis refers to going beyond simple word matching to analyze morphemes, syntactic structures, and inter-word relationships within a sentence to identify the intent of the utterance and its actual safety meaning. For example, from the sentence “Team Leader Park, please check the wire condition before crane operation, and Signalman, please ensure no one enters the work radius,” the device can extract “crane operation” as the work content or process, “wire” as the equipment used or the configuration of equipment subject to inspection, “prevent people from entering the work radius” as a safety instruction regarding access restrictions, and “Team Leader Park” and “Signalman” as information on participating workers or their roles. Additionally, expressions such as “could fall,” “could drop,” “could get caught,” and “risk of electric shock” can be recognized as expressions regarding risk factors and linked to risk classifications such as tipping, falling, entrapment, and electric shock. Here, meeting analysis information refers to a data set in which work-related information extracted from the meeting text is structured and stored.

[0048] The step of mapping the above meeting analysis information with at least one of a pre-established standard safety work procedure database, a risk assessment database, and a past accident case database to derive risk items and recommended safety measures corresponding to the above work content is a step that does not stop at merely recording the content actually mentioned in the meeting, but generates supplemented risk information and measure information through comparison and linkage with existing accumulated safety-related standard information.

[0049] Here, the standard safety work procedure database may refer to a database storing standard procedures for each work process, mandatory inspection items, standards for wearing protective equipment, equipment usage procedures, and prohibited items. The risk assessment database may refer to a database storing hazardous risk factors, risk levels, probability of occurrence, severity, existing reduction measures, and recommended preventive measures for each work process, and the historical accident case database may refer to a database storing accident cases, causes, consequences, conditions of occurrence, and subsequent corrective measures that occurred in the same or similar work processes.

[0050] The step of deriving risk items and recommended safety measures corresponding to the work content by mapping the above meeting analysis information to at least one of a pre-established standard safety work procedure database, a risk assessment database, and a past accident case database can be performed by querying each database using the work content, work process, equipment used, and risk expression included in the meeting analysis information as reference keys, determining whether there is a match or similarity between the queried reference information and the meeting content, supplementing missing risk items, and generating appropriate recommended safety measures. For example, even if only “exterior wall painting,” “use of aerial work platform,” and “caution against falling” are mentioned in the meeting text, the device can refer to the standard safety work procedure database and the risk assessment database to derive “falling” and “entrapment” as risk items and additionally derive recommended safety measures such as “verification of safety harness fastening,” “stationing of signalmen,” “securing of work platform outriggers,” and “control of access to the work radius.” Furthermore, if there is a history of accidents involving similar processes in the past accident case database, supplementary measures corresponding to the cause of the accident can be added and reflected in the document.

[0051] Here, "risk items" may refer to specific risk factors or inspection points requiring pre-work inspection, and "recommended safety measures" may refer to measures recommended to be performed or verified before work in response to the relevant risk items.

[0052] The step of generating a pre-work safety inspection document by automatically inputting content into a pre-configured TBM document template based on the aforementioned risk items, recommended safety measures, and participating worker information is a step of generating a final document by automatically editing and arranging the derived analysis results and supplemented safety information to fit the format of the pre-work safety inspection document.

[0053] Here, a TBM document template may refer to a predefined document form or electronic form that includes the work name, work date and time, work location, participating workers, equipment used, major risk items, key verification items, recommended safety measures, special notes, signature fields, and verification fields. Additionally, a pre-work safety inspection document may refer to a document prepared to share and verify the risk factors and preventive measures for the relevant work with participating workers before the start of work, and may be implemented as a paper document, electronic document, mobile document, checklist, or document subject to electronic approval.

[0054] The step of generating a pre-work safety inspection document by automatically inputting content into a pre-configured TBM document template based on the above-mentioned risk items, recommended safety measures, and participating worker information can be performed, for example, by placing work details and work processes included in the meeting analysis information in the work overview area at the top of the document, placing participating worker information in the attendee list area, and automatically entering risk items and recommended safety measures by aligning them with the risk checklist area. In this case, if multiple risk items exist, they can be entered by sorting them according to the order of risk level or the order of frequency of mention during the meeting, and recommended safety measures can be arranged in a row or column structure linked to each risk item. Additionally, the device can save the generated pre-work safety inspection document as a PDF, word processor document, electronic approval document, or structured data for a field inspection application, or transmit it to a manager terminal and a worker terminal. For example, the final generated document may be recorded in a format such as “Task Name: Ceiling Piping Installation,” “Equipment: Aerial Work Platform,” “Major Risks: Falls, Falling Objects, Insufficient Access Control,” “Major Actions: Safety Belt Verification, Work Radius Control, Equipment Leveling Check,” and “Participating Workers: 1 Team Leader, 3 Workers, 1 Signalman.” Accordingly, the work site can secure site-specific TBM documents immediately after the meeting, and can reduce omissions, delays, and quality variations caused by manual writing.

[0055] In addition, the step of deriving the meeting analysis information comprises: a step of primarily refining non-standard field terms, abbreviations, and colloquial expressions within the meeting text into standard safety terms by applying an industrial site-specific user dictionary, an expanded field abbreviation dictionary, and slang / dialect mapping metadata; a step of generating triplet data by dividing the primarily refined meeting text into sentence and clause units, and then performing Named Entity Recognition, Dependency Parsing, and Semantic Role Analysis to relation and structure words or phrases within the text into a triplet structure including Object (equipment used, tools, materials, or work objects), Action (work content, work sequence, or work process), and Condition (field environment, risk factors, or work conditions); and a step of generating work-action-risk association metadata by additionally tagging the work location, work time zone, type of equipment used, worker role, risk expression, safety instruction expression, and protective equipment-related expression for the triplet data. A step of generating a multidimensional embedding vector by inputting the above-mentioned work-behavior-risk association metadata into a pre-trained industrial safety specialized language model; a step of calculating the cosine similarity between the multidimensional embedding vector and the similar accident case sentence vector, risk situation description vector, and safety measure description vector included in a pre-set industrial safety standard corpus, respectively, to select a group of risk-associated sentences whose similarity score is greater than or equal to a pre-set threshold; a step of generating a group of potential risk factors corresponding to at least one of falling, entrapment, collision, electric shock, overturning, falling objects, fire, explosion, suffocation, collapse, and exposure to hazardous substances by analyzing the combination patterns of objects, behaviors, and states included in the group of risk-associated sentences; and a step of obtaining attendance verification data and work assignment information from a database.A step of extracting the name, affiliation, job title, and assigned task of a meeting participant using the meeting text, the speaker separation result derived from the voice preprocessing process, attendance confirmation data, and work assignment information, and generating participating worker information by obtaining qualifications, proficiency, recent work history, and safety training completion history from a database based on the name, affiliation, job title, and assigned task; a step of identifying imperative expressions, prohibitive expressions, confirmation request expressions, and attention-raising expressions from the work-behavior-risk association metadata, and classifying the identified imperative expressions, prohibitive expressions, confirmation request expressions, and attention-raising expressions into at least one safety instruction among instructions to wear protective equipment, instructions to prohibit access, instructions to comply with work sequence, instructions to inspect equipment, and instructions to check the environment; a step of obtaining external site environment information at the scheduled time of work and past accident history information of the target work process from a database. For each of the above-mentioned potential risk factor candidate groups, the method may include: a step of calculating a TBM reflection priority value for each of the above-mentioned potential risk factor candidate groups using a risk priority calculation model that applies individual worker proficiency, qualifications held, recent work history, safety training completion history, external site environment information at the scheduled time of work, and past accident history information of the target work process included in the above-mentioned participating worker information as weighting variables; a step of determining the item-by-item reflection order of caution phrases, items requiring verification, and preventive action phrases corresponding to the potential risk factor candidate groups based on the above-mentioned TBM reflection priority value; and a step of generating meeting analysis information including work content, work process, equipment used, risk factors, safety instructions, and participating worker information using the above-mentioned triplet data, potential risk factor candidate groups, safety instructions, and participating worker information.

[0056] The first refinement step is to convert non-standard field terms, abbreviations, and colloquial expressions included in the meeting text generated by the speech recognition model into a standard safety terminology system that allows for subsequent semantic analysis.

[0057] Here, the term "industrial site-specific user dictionary" may refer to a dictionary database that stores equipment names, process names, material names, hazard expressions, protective equipment names, role names, and work idioms that are repeatedly used in the field, along with their standard names. Additionally, the term "field abbreviation expansion dictionary" may refer to mapping information for expanding abbreviated expressions into full expressions that preserve their original meanings, and the term "slang and dialect mapping metadata" may refer to correspondence information for linking loanwords, slang, abbreviations, and dialect expressions customarily used in the field by industry or region with official names used in occupational safety and health-related laws or standard operating procedures.

[0058] In the first refinement step, for example, when sentences such as “I’m going to the exterior wall on a sky lift today,” “Don’t let anyone enter the hoist area,” “Watch out for the hook and make sure to wear your seatbelt,” and “I’m going to use a grinder on the scaffolding, so check your helmet,” exist in the meeting text, “Sky lift” can be normalized to “aerial work platform,” “hoist” to “hoisting machine or lifting equipment,” “belt” to “safety belt,” “scaffolding” to “scaffolding,” “grinder” to “grinding machine,” and “helmet” to “safety helmet,” respectively, by using the aforementioned industrial site-specific user dictionary, the aforementioned site abbreviation expansion dictionary, and the aforementioned slang and dialect mapping metadata. Additionally, colloquial expressions such as “Don’t let anyone enter” can be replaced with standard safety expressions such as “access restriction measures required” or “access control instructions.” Accordingly, the above meeting text can be converted into normalized text capable of subsequent named entity recognition, dependency parsing, and semantic role analysis while maintaining the speaker's original intent.

[0059] The step of generating triplet data is to divide the primary refined meeting text into sentence units and clause units, and then structure the words or phrases included in each divided section into three axes of object, action, and state to generate basic data for subsequent risk assessment.

[0060] Here, an object may refer to a noun phrase designating equipment, tools, materials, work objects, protective equipment, or specific facilities within the work site. An action may refer to a task execution movement or work procedure such as installation, dismantling, lifting, welding, cutting, inspection, movement, entry, fixing, fastening, verification, and waiting. A state may refer to an expression indicating the site environment, risk factors, working conditions, restrictions, attention levels, or temporal and spatial conditions. In this context, named entity recognition can perform the function of identifying equipment names, process names, worker names, location names, and protective equipment names; dependency parsing can perform the function of identifying modification relationships and dominance relationships within a sentence; and semantic role analysis can perform the function of interpreting who performs what under what conditions.

[0061] The step of generating triplet data can be structured such that, for example, when a sentence such as “Be careful when welding on the scaffolding because it is slippery due to the rain” is input, the device separates “scaffolding” or “welder” into objects, “welding work” into actions, and “rain” and “slippery” into states. Additionally, when a sentence such as “Team Leader Park should first check the wire wear condition before lifting the steel materials with the crane and ensure no one enters the work radius” is input, the device identifies “crane,” “steel materials,” “wire,” and “work radius” as objects, identifies “lifting,” “checking,” and “access control” as actions, and identifies “wear condition,” “pre-inspection required,” and “no entry” as states. Subsequently, the device analyzes the connection relationships between objects, actions, and states to generate multiple triplet data in the form of “wire-check-wear condition,” “work radius-access control-no entry,” and “crane-lifting-pre-inspection required.”

[0062] The step of generating work-behavior-risk association metadata involves adding tags to the triplet data—such as work location, work time zone, type of equipment used, worker role, risk expression, safety instruction expression, and protective equipment-related expression—to transform the simple triplet structure into expanded metadata from the perspective of on-site safety management. Here, "work location" may refer to physical location information such as work zones, floor numbers, equipment rooms, exterior walls, passageways, storage areas, confined spaces, or high-altitude work zones. "Work time zone" may refer to information indicating the temporal context of the work, such as morning work, afternoon work, night work, immediately after a shift change, before equipment operation, or specific time periods. "Type of equipment used" may refer to equipment classification information such as lifting equipment, cutting equipment, welding equipment, power tools, transport equipment, or equipment for attaching protective equipment, and "worker role" may refer to work supervisors, workers, safety managers, signalmen, equipment operators, inspectors, and supervisors.

[0063] The step of generating work-behavior-risk association metadata can, for example, tag “Zone 3 outer wall” as the work location, “8:00 AM” as the work time, “elevated work platform” as the type of equipment used, and “signalman” and “worker” as the worker roles, for a meeting text stating, “At 8:00 AM, painting work is performed on the outer wall of Zone 3 using an aerial work platform, the signalman controls access to the lower passage, and the worker must fasten their safety harness.” Additionally, “control access” can be tagged as a safety instruction expression, “must fasten safety harness” as an expression related to protective equipment and safety instruction expression, and “risk of falling” or “risk of falling objects” combined with “work at height” and “outer wall painting” can be tagged as a risk expression. Accordingly, in subsequent steps, risk can be assessed more precisely by using structured metadata that reflects the work context, role information, and risk expression together, rather than just the simple text itself.

[0064] The step of generating multidimensional embedding vectors is to input the aforementioned work-behavior-risk association metadata into a pre-trained industrial safety specialized language model to convert it into a multidimensional feature representation that quantifies semantic similarity, risk context, work procedure relationships, and safety instruction strength contained in the meeting text.

[0065] Here, an industrial safety-specialized language model may refer to a language model that has been additionally trained on information regarding work processes, equipment, hazard expressions, accident cases, preventive measures, names of protective equipment, and work procedures in the field of industrial safety, compared to a general language model. The above-mentioned multidimensional embedding vector may be generated corresponding to each sentence, clause, triplet, or metadata unit, and may be a semantic representation vector composed of a multi-dimensional array of real values.

[0066] The step of generating multidimensional embedding vectors can vectorize sentences such that, for example, the sentence “There is a risk of falling objects if the passageway below the work radius is not controlled from the top of the elevated work platform” and the sentence “A falling accident may occur if access to the lower area is not controlled during upper-level work” are judged to be similar at the semantic level, even if they differ at the character level. Conversely, “Checking equipment power before starting work” and “Securing a passageway during work” are both safety-related sentences, but because their semantic focus differs, they may have vector distributions in different directions. Such multidimensional embedding vectors can be used in the subsequent similarity calculation step to determine semantic proximity with similar accident cases or risk description sentences.

[0067] The step of selecting risk-associated sentence groups is to calculate the cosine similarity between the multidimensional embedding vector and the sentence vectors of similar accident cases, risk situation descriptions, and safety measure descriptions included in the pre-established industrial safety standard corpus, respectively, to select sentences or clauses within the meeting text that are highly relevant to actual risk judgment.

[0068] Here, an industrial safety standard corpus may refer to a set of standard sentences including industrial safety and health standard sentences, descriptions of standard operating procedures, descriptions of risk assessments, descriptions of accident cases, and standard precautionary measures.

[0069] A vector of sentences describing similar accident cases may be the result of vectorizing explanatory texts regarding the circumstances of past accidents; a vector of descriptions of hazardous situations may be the result of vectorizing sentences describing hazardous conditions that could occur in a specific work environment; and a vector of descriptions of safety measures may be the result of vectorizing standard phrases for mitigating specific risks. Cosine similarity is a mathematical technique that measures how similar the directionality of two sets of data is in space using the cosine value of the angle formed by two vectors; a higher value indicates a higher degree of semantic association.

[0070] The step of selecting risk-associated sentence groups involves calculating the cosine similarity for each sentence within the meeting text, such as “The scaffolding platform is slippery today,” “Keep people out of the equipment radius,” and “Check hook fastening before lifting.” If the similarity exceeds a preset threshold, the sentence can be selected as a risk-associated sentence group. In this case, “The scaffolding platform is slippery” may have high similarity to description vectors of risk situations related to slipping, overturning, or falling; “Keep people out of the equipment radius” may have high similarity to description vectors of safety measures related to access prohibition or collision prevention; and “Check hook fastening before lifting” may have high similarity to accident case sentences related to lifting equipment inspection. Accordingly, the device can incorporate only sentences with high actual risk association, rather than simply frequently occurring sentences, into the subsequent risk candidate group generation step.

[0071] The step of generating a group of potential risk factor candidates is to analyze the combination patterns of objects, actions, and states included in the group of risk-associated sentences to generate potential risk factor candidates corresponding to falling, entrapment, collision, electric shock, tipping, falling objects, fire, explosion, suffocation, collapse, and exposure to hazardous substances.

[0072] Here, the term "potential risk factor candidate group" refers to a set of candidates that are highly likely to lead to a specific risk based on object-action-state combination patterns, even if risk terms such as "falling" or "entrapment" were not explicitly and directly uttered during the meeting. For example, when "elevated work platform," "unidentified safety belt," "passing underneath," and "strong winds" appear together, potential risk factors related to falling and falling objects may be derived, and when "crane," "entry into work radius," and "no signalman stationed" appear together, potential risk factors related to collision or entrapment may be derived.

[0073] The step of generating a pool of potential risk factor candidates can be performed by comparing the objects, actions, and states included in each risk-associated sentence group with a predefined risk pattern rule set or by inputting them into a pre-trained risk classification model. For example, if the object is a “welding machine,” the action is “operation” or “work performance,” and the state is “insufficient ventilation” or “confined space,” risk factors related to fire, explosion, or suffocation may be generated as candidates. Additionally, if the object is a “switchboard” or “cable,” the action is “inspection” or “contact,” and the state is “power cutoff unconfirmed,” risk factors related to electric shock may be generated as candidates. The pool of potential risk factor candidates generated in this manner can be combined with various weighting variables in the subsequent priority calculation step to determine the priority for actual document reflection.

[0074] The step of obtaining attendance confirmation data and work placement information is a step of securing basic data for generating participating worker information by querying external management data from a database that indicates meeting attendance and the status of worker placement for the day.

[0075] Here, attendance verification data may refer to meeting attendance information generated by entry gate records, mobile attendance check records, manager confirmation input, time and attendance system records, or biometric authentication records. Additionally, work placement information may refer to placement information on a work assignment chart or work authorization system that includes the assigned process, input area, equipment, role, team composition, and working hours for each worker on that day. The above data may be obtained in conjunction with a site personnel database, a work management system, or an access control system.

[0076] The step of generating participating worker information is to extract the name, affiliation, job title, and assigned task of the meeting attendees using the meeting text, the speaker separation result derived from the voice preprocessing process, attendance confirmation data, and work assignment information, and to generate information on workers actually participating in the meeting or the day's work by obtaining qualifications, proficiency, recent work history, and safety training completion history from a database based on the extracted basic human information.

[0077] Here, the names, affiliations, duties, and assigned tasks of meeting participants can be identified through cross-referencing of names, self-introduction utterances, managerial instructions, attendance information, and work assignment information during the meeting. For example, if the meeting text contains remarks such as “Foreman Kim is in charge of scaffolding work in Zone 2 today, and Engineer Lee is going to operate the crane,” the device can identify “Foreman Kim” as the team leader or work manager and “Engineer Lee” as the equipment operator.

[0078] The step of generating participating worker information can be performed by querying the worker personnel database, qualification management database, training history database, and work performance database using the extracted name, affiliation, job title, and assigned task as key values ​​to obtain the qualifications, proficiency, recent work history, and safety training completion history of the worker.

[0079] Here, "qualifications held" may refer to qualifications required by statutory or internal standards, such as welding qualifications, crane operation qualifications, electrical work qualifications, working at height qualifications, or confined space work qualifications.

[0080] Proficiency may be a value defined based on the number of times the relevant process has been performed, the recent performance period, manager evaluation scores, or a pre-calculated competency index. Recent work history may refer to details of work related to similar processes or identical equipment performed within a certain period, and safety training completion history may refer to the history of completing regular safety training, new process training, equipment usage training, or special safety training. Accordingly, information on participating workers can be generated as structured data that goes beyond a simple list level and includes both quantitative and qualitative information usable for determining risk priorities.

[0081] The step of classifying into safety instructions is to identify imperative expressions, prohibitive expressions, confirmation request expressions, and attention-raising expressions from the above-mentioned work-behavior-risk association metadata, and to classify them into safety instructions such as instructions to wear protective equipment, instructions to prohibit access, instructions to follow work sequence, instructions to inspect equipment, and instructions to check the environment.

[0082] Here, imperative expressions may refer to phrases requiring the performance of specific actions, such as “confirm,” “secure,” “deploy,” or “inspect.” Prohibitory expressions may refer to phrases prohibiting specific actions, such as “do not enter,” “do not touch,” or “do not operate.” Confirmation request expressions may refer to phrases requesting a preliminary inspection or status check, such as “have you confirmed it?”, “look for any abnormalities,” or “check.” Additionally, warning expressions may refer to phrases that raise awareness of danger, such as “be careful,” “be cautious,” or “it is dangerous.”

[0083] In the step of classifying items into safety instructions, for instance, the sentence “Must fasten the chin strap of the safety helmet” can be identified as a commanding expression and classified as an instruction to wear protective equipment; the sentence “Never enter the radius of the crane” can be identified as a prohibitive expression and classified as an access restriction instruction; and the sentence “Check the wire condition first before lifting” can be classified as an equipment inspection instruction. Additionally, “Dismantle in order and do not stand people directly underneath” can be classified as an instruction to adhere to the work sequence and an access restriction instruction, while “Reconfirm whether to stop work if strong winds blow” can be classified as an environmental verification instruction. The safety instructions generated in this way can be utilized as core components that correspond to risk factors in the subsequent stage of generating analysis information for meetings.

[0084] The step of acquiring external site environment information and past accident history information of the target work process is a step of querying the database for accident history of processes identical or similar to the external environmental conditions at the scheduled time of work in order to calculate the priority of a group of potential risk factors.

[0085] Here, external site environment information may refer to information on external conditions affecting the work performance environment, such as weather data, temperature, humidity, wind speed, precipitation, wind chill, fine dust, and heatwave or cold wave advisories. Additionally, historical accident information may refer to historical data including the number of accidents, accident types, times of occurrence, causes, severity of injuries, frequency of recurrence, and details of related preventive measures that occurred in the past within the target work process or similar work processes. For example, if a strong wind forecast exists at the time of scheduled work at height, or if a history of repeated fall accidents in the same process is confirmed, such information may be reflected as an upward weighting factor when calculating risk priority.

[0086] The step of calculating the TBM reflection priority value is a step of quantifying the order of risk items that should be reflected preferentially within the document by using a risk priority calculation model that applies participating worker information, external site environment information, and past accident history information as weighting variables to each of the aforementioned potential risk factor candidate groups.

[0087] Here, the risk priority calculation model refers to a model that calculates the final priority for implementation by reflecting the individual worker's proficiency, qualifications, recent work history, safety training completion history, degree of external environmental deterioration, and accident history recurrence rate to the basic risk level for each potential risk factor. For example, even for the same fall risk candidate, if a worker with little experience working at heights is deployed, strong winds are forecast, and there have been numerous past fall accidents in the same process, the TBM implementation priority value for that risk candidate may be calculated as high. Conversely, if a skilled worker is deployed, the external environment is stable, and there is a low history of accidents in similar processes, a relatively low priority value may be calculated.

[0088] The step of calculating the priority value reflected in the TBM can be performed, for example, by assigning a basic risk score to each candidate group of potential risk factors, and then sequentially reflecting the inverse weighting of participating workers' proficiency, the penalty value for lack of qualification, the degree of lack of recent work history, the penalty value for failure to complete safety training, the external environment deterioration coefficient, and the coefficient of repeated past accidents to calculate the final priority value.

[0089] In addition, the device can further acquire pre-calculated work risk prediction information from a database to map risk factors by work process, unsafe worker behaviors, abnormal conditions in the field environment, signs of equipment malfunction, and the likelihood of accidents to a group of potential risk factor candidates. Subsequently, it calculates a risk prediction correction value reflecting the likelihood of accidents, the applicability of unsafe behaviors, abnormal field environmental conditions, and signs of equipment malfunction included in the aforementioned work risk prediction information, and applies this value to the priority base value with weights to correct the final TBM-reflected priority value. Accordingly, a priority can be calculated that incorporates not only risk judgments based solely on the meeting text but also the results of the pre-calculated field risk prediction.

[0090] The step of determining the order of reflection for each item is to decide the order in which cautionary statements, items requiring verification, and preventive action statements corresponding to the group of potential risk factors will be placed within the document, based on the aforementioned TBM reflection priority values. Here, cautionary statements may refer to phrases that directly warn or draw the worker's attention to a specific risk factor; items requiring verification may refer to items that need to be inspected before or during work; and preventive action statements may refer to specific action statements to prevent or mitigate the relevant risk. For example, if the risk of falling has the highest priority value, followed by the risk of falling objects and then the risk of insufficient access control, the device may arrange phrases such as "Verify safety belt fastening," "Control access to the lower area," and "Measures to prevent falling objects during upper-level work" in order of priority. Accordingly, in the subsequent document generation stage, the most critical risks and measures may be placed first at the top of the document or in a highly visible location.

[0091] The step of generating meeting analysis information is a step of synthesizing the triplet data, potential risk factor candidates, safety instructions, and participating worker information to generate final meeting analysis information including work content, work process, equipment used, risk factors, safety instructions, and participating worker information.

[0092] Here, the aforementioned triplet data provides the basic structure of text semantics, the aforementioned candidate group of potential risk factors provides interpretation results from a risk perspective, the aforementioned safety instructions provide the direction of preventive measures agreed upon during manager utterances or meetings, and the aforementioned participating worker information can provide information on personnel composition and worker capabilities. By integrating these elements into a single structured data object or a record composed of multiple fields, the device can ultimately generate meeting analysis information that can be used directly during the document generation stage.

[0093] The step of generating meeting analysis information can be performed by creating structured data in the form of, for example, “Work content: Exterior wall painting,” “Work process: Painting using an elevated work platform,” “Equipment used: Elevated work platform, safety harness,” “Risk factors: Falls, entrapment, falling objects,” “Safety instructions: Verification of safety harness fastening, placement of signalman, control of work radius,” and “Participating worker information: 1 team leader, 3 workers, 1 equipment operator.” At this time, the risk factors and safety instructions may be included in the meeting analysis information while maintaining the sorted order according to the previously calculated TBM reflection priority values. Accordingly, in the subsequent document generation step, the above meeting analysis information is automatically placed directly into the corresponding fields of the TBM document template, thereby enabling the consistent generation of site-specific pre-work safety inspection documents.

[0094] And, the step of calculating the priority value for TBM reflection for each of the aforementioned candidate groups of potential risk factors comprises: a step of obtaining pre-calculated work risk prediction information from a database; a step of extracting work process-specific risk factors, unsafe worker behaviors, abnormal site environmental conditions, signs of equipment abnormalities, and the probability of accident occurrence included in the said work risk prediction information; a step of mapping the said work risk prediction information and the candidate groups of potential risk factors based on commonalities among work processes, work locations, equipment used, work time zones, and participating workers; a step of calculating a risk prediction correction value that reflects the probability of accident occurrence, whether unsafe behaviors apply, whether abnormal site environmental conditions apply, and whether signs of equipment abnormalities apply, based on the mapping results for each of the said candidate groups of potential risk factors; and a step of calculating a basic priority value for each of the candidate groups of potential risk factors based on the individual worker's proficiency, qualifications held, recent work history, safety training completion history, external site environment information at the scheduled time of work, and past accident history information of the target work process included in the said participating worker information. and may include the step of determining TBM reflection priority values ​​for each potential risk factor candidate group by applying the above risk prediction correction value to the priority base value with weights.

[0095] The step of obtaining pre-calculated work risk prediction information from the database is a step for utilizing dynamic risk data from the site, which is pre-calculated and stored by a separate safety control system, an artificial intelligence analysis server, or an integrated safety management platform, in addition to the group of potential risk factors derived from the meeting text itself, for subsequent priority calculation.

[0096] Here, work risk prediction information refers not merely to information relying on historical statistics or regulatory documents, but to structured data derived by integrating and analyzing on-site IoT sensors, intelligent closed-circuit television, equipment operation logs, work plan information, work permit information, and safety-related document information in a multimodal format; this data may include risk factors by work process, unsafe worker behaviors, abnormal conditions of the on-site environment, signs of equipment malfunction, and the likelihood of accidents. Furthermore, the aforementioned database may be a repository within an on-site server, a cloud-based risk prediction repository, a history database of an integrated safety management platform, or a prediction result repository of an externally linked system.

[0097] The step of obtaining pre-calculated work risk prediction information from a database can be performed by searching for said work risk prediction information using, for example, the daily work schedule, work site identification information, work process code, work zone information, equipment identification information, and meeting time information as lookup keys, and by selecting and retrieving the most recently generated prediction result or a prediction result with a high degree of agreement with the current work. In this case, if multiple prediction results exist, the device can determine the priority application target based on the generation time, prediction validity period, work agreement, or risk level. Accordingly, the site risk prediction result already calculated during or immediately before the meeting can be combined with the meeting analysis flow.

[0098] The step of extracting risk factors by work process, unsafe worker behaviors, abnormal conditions of the site environment, signs of equipment abnormalities, and the possibility of accidents included in the above-mentioned work risk prediction information is a step of separating and extracting, by detailed item, the key elements directly necessary for reflection in the TBM document among the multiple prediction items included in the above-mentioned work risk prediction information.

[0099] Here, risk factors by work process may refer to potential risks inherent in a specific process itself or types of risks highly likely to occur in a specific process, and unsafe worker behavior may refer to behavioral patterns that increase the likelihood of accidents, such as not wearing protective equipment, entering hazardous areas, violating work sequences, approaching equipment in close proximity, or careless actions. Additionally, abnormal site environmental conditions may refer to conditions where the working environment has deteriorated abnormally, such as strong winds, rain, high or low temperatures, low light levels, slippery conditions, insufficient ventilation, or increased concentration of hazardous gases, and signs of equipment abnormality may refer to signs where the equipment condition deviates from the normal range, such as abnormal vibration, overheating, abnormal current, abnormal pressure, the occurrence of warning alarms, or repetitive malfunctions.

[0100] The step of extracting risk factors by work process, unsafe worker behavior, abnormal conditions of the site environment, signs of equipment abnormalities, and the probability of accidents included in the above-mentioned work risk prediction information can, for example, if the above-mentioned work risk prediction information includes items such as “high probability of falling during high-altitude work processes,” “frequent worker access within the lifting radius of Zone 2,” “increased risk of exterior wall work due to rising wind speed,” “increased hoisting machine vibration and wire tension imbalance,” and “78 percent probability of accident,” the device can separate and extract these into risk factors by work process, unsafe worker behavior, abnormal conditions of the site environment, signs of equipment abnormalities, and the probability of accidents, respectively. At this time, the probability of accidents may be in the form of a percentage, probability score, risk index, or grade value, and may be used as a normalized numerical value in the subsequent correction value calculation step.

[0101] The step of mapping the work risk prediction information and the group of potential risk factors based on commonalities among work processes, work locations, equipment used, work time zones, and participating workers is a step of forming a correspondence relationship between the group of potential risk factors derived from the meeting analysis results and the work risk prediction information obtained from the prior risk prediction system.

[0102] Here, commonality may refer to semantic or identifier-based consistency, including whether the process is the same, the work location is the same, the equipment used is the same, the time zone is the same or adjacent, or the same participating workers or role groups are the same. In other words, it may be a procedure that determines whether the risk candidates mentioned in the meeting and the risk items indicated by the prediction system share the same work context and links them to each other.

[0103] The step of mapping the work risk prediction information and the group of potential risk factors based on commonalities among work processes, work locations, equipment used, work time zones, and participating workers can, for example, map the two to the same risk series based on work process similarity, location consistency, equipment consistency, time zone consistency, and participating worker consistency, if, for instance, the meeting analysis results generate "steel frame assembly work in Area A," "use of aerial work platforms," ​​and "risk of falling" as the group of potential risk factors, and the prior work risk prediction information derives "steel frame assembly process in Area A," "operation of aerial work platforms," ​​"detection of unfastened safety harness," and "high probability of falling accidents under strong wind conditions." Additionally, even if the meeting analysis results derive "risk of entering within radius during crane lifting" and the prior prediction information contains "frequent behavior of approaching the radius of lifting equipment," they can be mapped to risks related to collision or entrapment for the same work process and the same equipment used. The results mapped in this manner can be used as input for calculating subsequent risk prediction correction values.

[0104] Based on the mapping results for each of the aforementioned potential risk factor candidate groups, the step of calculating a risk prediction correction value that reflects the probability of accident occurrence, whether unsafe behavior applies, whether the site environment is in an abnormal state, and whether there are signs of equipment abnormality included in the work risk prediction information is a step of reflecting the risk judgment results of the pre-prediction system in the form of correction coefficients to the meeting-based analysis results.

[0105] Here, the risk prediction correction value may refer to a correction value used to adjust the base priority of a group of potential risk factor candidates upward or downward. The aforementioned risk prediction correction value may increase as the probability of an accident increases, and may be further weighted if unsafe worker behavior, abnormal site environmental conditions, and signs of equipment abnormality coexist within the same candidate group.

[0106] Based on the mapping results for each of the aforementioned potential risk factor candidate groups, the step of calculating a risk prediction correction value that reflects the probability of accident occurrence, whether unsafe behavior applies, whether abnormal site environmental conditions apply, and whether signs of equipment abnormality apply, included in the work risk prediction information, is, for example, when the potential risk factor candidate group is fall risk and the corresponding prediction information indicates a high probability of accident occurrence, true for unsafe behavior, abnormal site environmental conditions due to strong winds, and signs of equipment abnormality in the hydraulic system of the elevated work platform, the device can calculate a high risk prediction correction value by sequentially reflecting the unsafe behavior weight, the environmental abnormality weight, and the equipment abnormality weight to the basic accident probability score. For example, with the basic correction coefficient set to 1.0, if the probability of accident occurrence is evaluated as very high and equipment abnormality and unsafe behavior are detected simultaneously, a correction value of 1.5 or 2.0 can be calculated. Conversely, if the probability of accident occurrence is low and there are no additional abnormal factors, a relatively low correction value may be calculated. In this case, whether it constitutes unsafe behavior, whether it constitutes an abnormal state of the site environment, and whether it constitutes an abnormal sign of equipment can be reflected in the form of binary values, multi-level risk level values, or normalized probability values.

[0107] The step of calculating basic priority values ​​for each group of potential risk factors based on the individual worker's proficiency, qualifications, recent work history, safety training completion history, external site environment information at the scheduled time of work, and past accident history information of the target work process included in the above participating worker information is a step of calculating the basic priority of each risk candidate based on worker conditions and external environment conditions directly linked to the meeting results in the state prior to real-time prediction correction.

[0108] Here, the priority baseline value may be a base score indicating the weight with which a group of potential risk factors should be reflected in the TBM document in the pre-calibration state of the advance prediction system. The aforementioned proficiency may be the level of experience in performing the relevant process or a manager's evaluation value; qualifications may be statutory qualifications or internal certifications; and recent work history may be the number of times similar tasks were performed or the duration of such work within a certain period. Additionally, safety training completion history may indicate whether regular safety training, special safety training, or new process training has been completed, as well as the date of the most recent completion.

[0109] The step of calculating a priority base value for each candidate group of potential risk factors based on the individual worker's proficiency, qualifications, recent work history, safety training completion history, external site environment information at the scheduled time of work, and past accident history information of the target work process included in the above participating worker information is such that, for example, regarding a candidate group of fall risk, if it includes new workers with little experience in working at heights, lacks relevant qualifications, has a low number of recent similar process performances, lacks records of completing special safety training for working at heights, has a strong wind forecast at the scheduled time of work, and has a high frequency of past fall accidents in the same process, the device can calculate a high priority base value by reflecting these factors as upward weighting factors. For example, a priority base value of 75 points can be calculated by reflecting penalties for insufficient proficiency, lack of qualifications, lack of work history, lack of training completion, external environment deterioration coefficient, and accident history repetition coefficient to a basic risk score of 50 points. Conversely, if skilled workers are deployed and adequately qualified, have sufficient recent work history and good training completion status, the external environment is stable, and there is a low history of past accidents, a relatively low priority baseline value may be calculated.

[0110] The step of determining the TBM reflection priority value for each potential risk factor candidate group by applying the above risk prediction correction value to the priority base value with weights is a step of determining the final TBM document reflection priority by integrating the priority base value calculated from meeting-based risk judgment and the risk prediction correction value derived from the prior risk prediction system.

[0111] Here, the TBM reflection priority value may be a final numerical value indicating the order and intensity with which each candidate group of potential risk factors should be reflected within the safety inspection document prior to final work. The above-mentioned weighting application may be performed using simple addition, ratio weighting, grade correction, or threshold-based upward adjustment methods, and in one embodiment, it may be implemented by multiplying the priority base value by a risk prediction correction value.

[0112] The step of determining the TBM reflection priority value for each potential risk factor candidate group by applying the above risk prediction correction value to the priority base value by weight is, for example, when the priority base value of a fall risk candidate group is calculated as 75 points, and the risk prediction correction value is calculated as 1.2 because the probability of an accident is high and unsafe behavior and signs of equipment abnormality are simultaneously confirmed through CCTV video and sensor analysis, the device can derive a final TBM reflection priority value of 90 points by multiplying 75 points by 1.2. Conversely, if the priority base value is at an intermediate level but the actual probability of an accident is low based on prior prediction results and there are no additional signs of abnormality, a small correction value is applied, which may limit the increase in the final priority value.

[0113] Accordingly, a dynamic priority can be calculated that reflects not only risk assessments based solely on the meeting text itself but also pre-calculated field risk prediction results, and in subsequent steps, the placement order of cautionary statements, items requiring verification, and preventive action statements can be determined more rationally based on the priority values ​​reflected in the TBM.

[0114] The above-described embodiment and the following embodiment may be implemented by combining some or all of the components.

[0115] According to one embodiment, an integrated safety management method for AI-based pre-identification of work risks and calculation of dynamic safety scores, wherein a device includes a processor, memory, a communication module, and a non-transient storage medium, and the processor executes a program stored in the non-transient storage medium, comprises: a step of acquiring a plurality of work-related data including IoT environment sensor data, CCTV video data, work plan data, work permit data, meeting audio data, safety inspection documents, and equipment log data from a work site; a step of generating integrated input data for analysis by performing preprocessing on the work-related data including data format unification, noise removal, missing value correction, time synchronization, and normalization; a step of deriving work risk prediction information including risk factors by work process, worker unsafe behavior, abnormal conditions of the site environment, signs of equipment abnormality, and the possibility of accident occurrence by performing multimodal fusion analysis on the integrated input data for analysis using a pre-trained artificial intelligence model; and a step of calculating worker behavior indicators, site condition indicators, and organizational management indicators based on the work risk prediction information. The present invention provides an AI-based integrated safety management method for pre-identifying work risks and calculating dynamic safety scores, comprising the step of calculating an integrated safety score by applying pre-set weights to the worker behavior indicators, site condition indicators, and organizational management indicators, and assigning a safety grade corresponding to the integrated safety score.

[0116] The step of acquiring multiple work-related data from the work site is a step in which a device collects and links heterogeneous data generated inside and outside the site to analyze the safety status of the day's work from various angles.

[0117] The step of acquiring multiple work-related data from the work site may be a step of collecting IoT environment sensor data, CCTV video data, work plan data, work permit data, meeting audio data, safety inspection documents, and equipment log data.

[0118] Here, IoT environmental sensor data may refer to sensor measurement data indicating the state of the working environment, such as temperature, humidity, wind speed, hazardous gas concentration, fine dust levels, illuminance, noise levels, vibration values, or atmospheric pressure values. CCTV video data may refer to real-time streaming video or recorded video files received from camera equipment filming work areas, movement pathways, elevated work zones, lifting radii, or the surroundings of facilities. Additionally, work plan data, work permit data, meeting audio data, and safety inspection documents may refer to unstructured data in text or audio form containing details of the processes to be performed on the day, approval records, remarks exchanged during pre-work toolbox meetings, and items inspected by managers. Furthermore, equipment log data may refer to records of operating status, warning alarm occurrences, error codes, current values, temperature values, motor status values, or vibration level fluctuations of heavy equipment and critical facilities, such as cranes, forklifts, press machines, hoists, elevated work platforms, welding machines, or switchboards.

[0119] The step of acquiring multiple work-related data from a work site can be performed, for example, by receiving sensor packets through a field gateway, reading CCTV footage of a specific time interval from a video storage server, receiving the daily work plan from a work management system, querying approved work permits from a permit management system, collecting safety meeting recording files from a conference terminal, and retrieving equipment log files from an equipment management server. At this time, each data set may be assigned work site identification information, work zone identification information, work process identification information, equipment identification information, collection time, and data source information. Accordingly, the device can link raw data of different formats within the same work site and timeframe context.

[0120] The step of generating integrated input data for analysis by performing preprocessing on the data related to the above-mentioned task is a step of organizing the collected heterogeneous data into a common format that enables subsequent artificial intelligence analysis.

[0121] Here, data format unification refers to a process that unifies different storage formats by performing time-series standardization of sensor data, frame-unit reconstruction of image data, text extraction of document data, conversion of voice data to standard voice file formats, and field normalization of log data. Noise removal refers to a process that removes or mitigates sensor outliers, image shake, voice noise, document recognition errors, or unnecessary duplicate events within logs. Missing value correction refers to a process that supplements missing sensor measurements, uncollected frame intervals, partially empty document fields, or corrupted log entries with values ​​from adjacent time points, averages, medians, or interpolated values. Time synchronization refers to a process that aligns sensor collection times, image frame times, log occurrence times, meeting times, and scheduled task start times along the same reference time axis. Normalization refers to a process that converts values ​​with different units into comparable numerical values ​​within a certain range to adjust the data so that the size of specific data does not excessively affect the analysis results.

[0122] The step of generating integrated input data for analysis by performing preprocessing on the aforementioned work-related data can, for example, arrange wind speed and hazardous gas concentration values ​​among sensor data in chronological order and remove outliers; sample CCTV video data at specified frame intervals and resize it to a resolution where workers and equipment are identifiable; separate work plans and work permits into text fields; convert meeting audio data into text using a speech recognition model; and reconstruct equipment log data by error code, warning occurrence time, and measurement value fields. In particular, time synchronization may refer to aligning the frame times of videos collected from different systems, the measurement times of sensors, and the log occurrence times of equipment based on the same timestamp, and normalization may refer to scaling the ranges of sensor values ​​with different ranges and units to values ​​between zero and one to ensure the stability of subsequent analysis. Subsequently, the device can generate integrated input data for analysis with a mutually corresponding structure by assigning the same work zone code and time window to each data. For example, for exterior wall work from 8:00 AM to 8:30 AM, wind speed data for the relevant section, work radius video, exterior wall painting plan, high-altitude work permit conditions, meeting remarks, and high-altitude work platform hydraulic logs can be integrated into a single input set.

[0123] The step of deriving work risk prediction information by performing multimodal fusion analysis using an artificial intelligence model on the aforementioned integrated input data for analysis is a step of predicting complex risks that are difficult to identify with single data alone by analyzing time series data, image data, text data, and log data together.

[0124] Here, multimodal fusion analysis refers to an analysis method that combines feature values ​​or latent expressions extracted from different data forms to generate a single comprehensive risk assessment result. Work risk prediction information refers to structured risk analysis results that include risk factors by work process, unsafe worker behaviors, abnormal site environmental conditions, signs of equipment anomalies, and the likelihood of accidents. Risk factors by work process may refer to inherent risks associated with tasks to be performed on the day, such as steel lifting, work at heights, hot work, or confined space work, identified by analyzing text data. Unsafe worker behaviors may refer to human-induced risk-causing actions identified by analyzing video data, such as not wearing a safety helmet, walking while using a smartphone, unauthorized entry into hazardous areas, or failure to maintain a safe distance from equipment. Furthermore, abnormal site environmental conditions and signs of equipment anomalies may refer to factors based on sensor and log data, such as hazardous gases approaching the lower explosion limit, the risk of heatstroke due to rising temperatures, abnormal equipment vibration patterns, or the occurrence of repetitive warning alarms.

[0125] The step of deriving work risk prediction information by performing multimodal fusion analysis using an artificial intelligence model on the integrated input data for analysis described above can be performed, for example, by detecting a surge in hazardous gas concentration, exceeding temperature and humidity thresholds, or abnormal wind speed from sensor data; detecting non-wearing of protective equipment, unsafe movement paths, entry into hazardous areas, or failure to maintain a safe distance from equipment from CCTV footage; extracting potential risks inherent in the work process of the day from work plans, permits, meeting texts, and safety inspection documents; and detecting abnormal vibration, overheating, repeated warnings, or abnormal operation patterns from equipment log data. Subsequently, the device can combine the multiple analysis results to generate work risk prediction information such as, for example, “increased risk of falls and falling objects in exterior wall high-altitude work processes,” “frequent worker access within the lifting radius,” “deterioration of the site environment due to strong winds,” “persistent abnormal vibration of the hoisting machine,” and “82 percent probability of an accident.” Here, the probability of an accident occurring may be a numerical value representing the probability of a type of serious accident, such as a fall, explosion, entrapment, or collision, occurring in a specific time interval or a specific work process.

[0126] The step of calculating worker behavior indicators, site condition indicators, and organizational management indicators based on the aforementioned work risk prediction information is a step of converting the predicted risk information into manageable indicator forms at the worker, site, and organizational levels. Here, worker behavior indicators may refer to indicators representing the level of unsafe behavior by individual workers or role groups, compliance with protective equipment, propensity to approach hazardous areas, or the degree of repeated violations. Site condition indicators may refer to indicators representing the degree of environmental deterioration in the work area, instability of conditions by process, concentration of equipment abnormalities, or the degree of simultaneous occurrence of multiple risks. Organizational management indicators may refer to indicators reflecting the operational soundness of the management system, such as the repetitiveness of the same type of accident, the degree of simultaneous exposure of multiple workers, the simultaneous high-risk status of multiple processes, risk diffusion, the burden of management intervention, the timely approval rate of work permits, or the completeness of meeting and safety inspection documents.

[0127] The step of calculating worker behavior indicators, site condition indicators, and organizational management indicators based on the aforementioned work risk prediction information can, for example, calculate worker behavior indicators by normalizing and assigning weights when three instances of not wearing a safety helmet, two instances of approaching the lifting radius, four instances of failing to maintain a safety distance, and a cumulative duration of 120 seconds of abnormal posture are detected for a specific worker. Additionally, site condition indicators can be calculated by combining the detection of increased hazardous gas concentration, decreased illumination, strong winds, abnormal equipment vibration, and video-based congestion conditions in a specific work area. Furthermore, if fall and collision risks increase simultaneously across multiple work areas within the same site and multiple workers are exposed at the same time, organizational management indicators can be calculated by evaluating this as a complex risk at the organizational management level. Unlike a simple list of events, the results quantified in this way can serve as a means to intuitively determine management priorities.

[0128] The step of calculating an integrated safety score by applying pre-set weights to the above-mentioned worker behavior indicators, site condition indicators, and organizational management indicators, and assigning a safety grade corresponding to the integrated safety score, is a step of converting different levels of safety risk into a single comprehensive score.

[0129] Here, pre-set weights may refer to reflection ratios that are set in advance or periodically adjusted based on site characteristics, process risk, workplace operation policies, equipment importance, or sensitivity to serious accidents. The integrated safety score may be a comprehensive score representing the overall safety level of the current work site by integrating worker behavior indicators, site condition indicators, and organizational management indicators into a single numerical value. The safety rating may refer to evaluation results classified into categories such as Safe, Caution, Warning, or Danger based on the aforementioned integrated safety score.

[0130] The step of calculating an integrated safety score by applying pre-set weights to the above-mentioned worker behavior indicators, site condition indicators, and organizational management indicators, and assigning a safety grade corresponding to the integrated safety score, can be performed, for example, by assigning a weight of 30 percent to the worker behavior indicators, 40 percent to the site condition indicators, and 30 percent to the organizational management indicators, and then generating an integrated safety score based on a maximum of 100 points by weighting and summing each indicator. As another example, depending on the characteristics of the work site, a weight of 40 percent to the worker behavior indicators, 40 percent to the site condition indicators, and 20 percent to the organizational management indicators may be assigned. For example, if the above-mentioned worker behavior indicators are 80 points, the above-mentioned site condition indicators are 70 points, and the above-mentioned organizational management indicators are 90 points, and the weights are 40 percent, 40 percent, and 20 percent, respectively, the device can calculate an integrated safety score of 78 points by summing 32 points, 28 points, and 18 points. Subsequently, safety grades can be assigned according to predefined criteria, such as safety for 80 points or higher, caution for 60 points or higher but less than 80 points, warning for 40 points or higher but less than 60 points, and danger for less than 40 points. Accordingly, managers or orderers can quickly identify the complex safety status of the site with a single score and grade, and more efficiently determine whether to prioritize intervention for high-risk workers, high-risk work zones, and high-risk processes.

[0131] In addition, the step of deriving the above-mentioned work risk prediction information comprises: a step of deriving abnormal field environmental conditions and signs of equipment abnormalities, and extracting time-series latent vectors by applying a prediction model based on a Long Short-Term Memory (LSTM) algorithm to time-series analysis of patterns such as surges in hazardous gas concentration, exceedance of temperature and humidity thresholds, and abnormal equipment vibration within a preset time window for time-series stream data including IoT environment sensor data and equipment log data among the above-mentioned integrated input data for analysis; and a step of identifying unsafe worker behaviors and extracting spatial behavioral latent vectors by applying a vision AI model, which combines object detection and 3D skeleton estimation algorithms, to CCTV video data among the above-mentioned integrated input data for analysis to analyze spatiotemporal behavioral patterns including the worker's non-wearing of personal protective equipment (PPE), unsafe movement paths, entry into hazardous areas, and failure to maintain a safe distance from heavy equipment. A step of extracting risk factors for each work process, including inherent potential risks of the work process on the day, and extracting text embedding vectors by performing named entity recognition (NER) and contextual analysis using an industrial safety-specific pre-trained language model (Domain-Specific Transformer) on text data converted from work plan data, work permit data, and meeting voice data, and safety inspection documents among the integrated input data for analysis; a step of simultaneously inputting the time-series potential vector, spatial behavioral potential vector, and text embedding vector into a Cross-Attention network layer to generate a single multi-modal fused vector reflecting the mutual correlation weights between heterogeneous data;The method may include the step of inputting the above multimodal fusion vector into a gradient boosting-based ensemble classifier (XGBoost) to calculate the probability of accident occurrence by type of major accident, including falls, entrapment, collision, electric shock, fire, explosion, and suffocation, at a specific point in time at a target work site as a probability index in percentage units; and the step of generating work risk prediction information by combining the said probability of accident occurrence with risk factors, unsafe behaviors, abnormal conditions of the site environment, and signs of equipment abnormality.

[0132] The step of deriving abnormal field environmental conditions and signs of equipment abnormalities by applying a prediction model based on a long-term short-term memory algorithm to time-series stream data, and extracting time-series latent vectors, is a step of interpreting changes in the environment and equipment status of the work site in a time-series manner using numerical data continuously collected in chronological order among the integrated input data for analysis.

[0133] Here, time-series stream data refers to data having a continuous flow, such as measurements from environmental sensors and operation log records of equipment that are continuously collected at regular time intervals, and may include, for example, hazardous gas concentration, temperature, humidity, wind speed, dust concentration, equipment current value, motor temperature, vibration value, pressure value, whether a warning alarm occurred, and equipment operating status code. Additionally, a prediction model based on a long-term short-term memory algorithm refers to a recurrent neural network-based model suitable for predicting anomalies in current data or short-term future states by reflecting both data patterns accumulated over a long period of time and change patterns at a recent point in time. In this case, a time window refers to a specific time interval within the entire time-series data that serves as the unit of analysis, and may be set, for example, as a data block in units of the last 10 minutes, the last 30 minutes, or the last 1 hour.

[0134] The step of deriving abnormal field environmental conditions and signs of equipment abnormalities, and extracting time-series latent vectors by applying a prediction model based on long-term and short-term memory algorithms to time-series stream data, can be performed, for example, by analyzing the rate of change in hazardous gas concentration within a recent time window, the upward trend of temperature and humidity, wind speed variability, changes in the periodicity of equipment vibration, and the cumulative rise pattern of motor temperature. In this case, a surge in hazardous gas concentration can be determined when the concentration value increases above a preset reference slope, an exceedance of the temperature and humidity threshold can be determined when the allowable range for each work process is exceeded, and an abnormal equipment vibration pattern can be determined when an increase in amplitude, frequency deviation, repetitive vibration, or instantaneous shock pattern occurs by comparing it with the vibration distribution of the normal operating section. For example, if the oxygen concentration decreases in a confined space work area while the hazardous gas concentration surges rapidly within a short period, and simultaneously the current value of the ventilation equipment drops abnormally, the device can identify this as a state where both an abnormal field environmental condition and signs of equipment abnormalities exist simultaneously. Here, abnormal site environmental conditions and signs of equipment abnormalities may refer to physical and mechanical risk factors such as hazardous gas leakage, excessive heat, lack of ventilation, abnormal motor vibration, or current instability.

[0135] The step of extracting a time series latent vector is to extract a feature representation immediately prior to the final output or a hidden state formed within the prediction model based on the long-term short-term memory algorithm, and to generate a vector form that compresses the temporal characteristics and abnormal patterns of the time series data.

[0136] Here, a time-series latent vector may be a multi-dimensional numerical representation that reflects trends in hazardous gas concentration, changes in temperature and humidity, wind speed fluctuations, equipment vibration abnormalities, and the cumulative pattern of warning alarms. For example, even if two pieces of equipment have the same average vibration value, if one shows a gradual rising pattern and the other shows an intermittent shock pattern, the time-series latent vector can reflect this difference as vector values ​​of different directions or magnitudes. Accordingly, in the subsequent multimodal fusion step, high-dimensional feature values ​​reflecting temporal meaning can be used instead of simple raw sensor values.

[0137] The step of identifying unsafe worker behaviors and extracting spatial behavior potential vectors by applying a vision artificial intelligence model that combines object detection and 3D skeleton estimation algorithms to CCTV video data is a step of analyzing the worker's movements, postures, positional relationships, and spatial interactions with equipment based on video data among the integrated input data for analysis.

[0138] Here, an object detection algorithm may refer to a technology that identifies objects within a video frame, such as workers, safety helmets, safety belts, safety vests, cranes, forklifts, aerial work platforms, lifted objects, switchboards, or danger zone boundaries, in terms of location and category. Additionally, a 3D skeleton estimation algorithm may refer to a technology that interprets work postures and movement patterns by estimating the joint positions of a worker, such as the head, shoulders, elbows, wrists, pelvis, knees, and ankles, in the form of three-dimensional coordinates. Furthermore, a vision artificial intelligence model may refer to a video analysis model that determines the risk of a worker's behavior by combining the object detection results and the 3D skeleton estimation results.

[0139] The step of identifying unsafe worker behaviors and extracting spatial behavior potential vectors by applying a vision AI model that combines object detection and 3D skeleton estimation algorithms to CCTV video data can be performed, for example, by determining as unsafe behavior cases where a specific worker enters a work area without wearing a safety helmet, repeatedly approaches within the lifting radius, leans their upper body excessively outside the railing of an elevated work platform, or moves without securing a safe distance from heavy equipment.

[0140] Here, the state of not wearing protective equipment can be determined by comparing whether the worker's head area coincides with a safety helmet object, unsafe movement paths can be determined by comparing the boundary of a danger zone with the worker's movement trajectory, and entry into a danger zone can be determined based on whether the worker's location coordinates intersect with a pre-set prohibited area polygon. Additionally, failure to maintain a safe distance from heavy equipment can be determined when the distance between the worker and a forklift, crane, aerial work platform, or rotating equipment is less than a pre-set threshold distance. For example, if a series of movement data is identified in which a worker without a safety helmet slowly approaches the working radius of an operating excavator, the device can identify this as a high-risk unsafe behavior combining the failure to wear protective equipment and approaching heavy equipment.

[0141] The step of extracting spatial behavior potential vectors is a step of compressing and expressing the worker posture, movement speed, movement direction, distance relationship with equipment, degree of overlap with danger zones, and whether protective equipment is worn, analyzed by the vision AI model, into a high-dimensional vector form.

[0142] Here, a spatial behavior potential vector refers to a feature representation that quantifies a worker's spatiotemporal behavioral patterns. Even for the same entry into a hazardous area, it can be expressed in different vector forms depending on the frequency of entry, duration of stay, distance from equipment, and the worker's role. For example, while a worker simply crossing a passageway and staying directly beneath a load being lifted for an extended period may both technically be considered approaches to the lifting radius, the spatial behavior potential vector can be generated to reflect the risk level of the latter more significantly. Accordingly, in the subsequent multimodal fusion stage, three-dimensional feature values ​​reflecting location, attitude, distance, and time information can be used instead of simple behavior labels.

[0143] The step of extracting risk factors by work process and extracting text embedding vectors by performing named entity recognition and contextual analysis on text data converted from work plan data, work permit data, and meeting voice data, as well as safety inspection documents using an industrial safety-specific pre-trained language model, is a step of interpreting potential risks inherent in the work process itself on the day using text series data among the integrated input data for analysis.

[0144] Here, an industrial safety-specialized pre-trained language model refers to a model based on a Transformer architecture that enhances the performance of understanding safety domain context by additionally training on industrial safety and health laws, field work guidelines, historical accident records, risk assessment documents, TBM logs, and work permit documents. Furthermore, named entity recognition refers to the function of identifying semantic units within text, such as work process names, equipment names, material names, location names, protective equipment names, hazardous material names, role names, or worker names, while contextual analysis refers to the function of deriving potential risks by interpreting relationships between words, sentence intent, and the narrative context of risk expressions. In this context, the inherent potential risk of the day's work process refers to risk factors fundamentally hidden within the process itself to be performed on that day, which are identified by analyzing expressions such as confined spaces, suffocation risks, hot work, work at heights, or lifting operations listed in work plans or inspection documents, along with their surrounding context.

[0145] The step of extracting risk factors by work process and text embedding vectors by performing named entity recognition and contextual analysis on text data converted from work plan data, work permit data, and meeting audio data, as well as safety inspection documents, using an industrial safety-specialized pre-trained language model, allows the device to recognize the installation of exterior wall panels as a complex process combining work at height and lifting operations in the work plan, the work permit as requiring simultaneous high-altitude and lifting operations in the meeting text as caution regarding falling objects due to the possibility of strong winds, and the safety inspection document as requiring safety harnesses, signalmen, and lower-level control. Furthermore, by interpreting expressions such as strong winds, falling objects, and lower-level control as contexts for risks and measures, the device can extract risk factors by work process, such as falls, falling objects, and collisions. Additionally, if the process itself is inherently linked to specific serious accidents, such as confined space cleaning, fire work, electrical inspection, and heavy lifting operations, the language model can recognize the inherent risks of the corresponding process with a high priority.

[0146] The step of extracting text embedding vectors is to convert the text meaning interpreted by the industrial safety-specialized pre-trained language model into a multi-dimensional numerical vector to generate a feature representation that compresses the risk context and work procedure relationship of the text data.

[0147] Here, a text embedding vector refers to a vector that quantifies semantic information contained in work plans, permit conditions, meeting remarks, and inspection documents into a continuous real-valued space. For example, the need to control the lifting radius and the prohibition of access to lower passages can be represented as closely related vectors because, although their sentence expressions differ, they are semantically similar; similarly, gas measurement before hot work and the prohibition of welding in confined spaces can be represented as vectors with high relevance in the context of fire and explosion. Accordingly, in the subsequent cross-attention stage, the text semantics can be organized into a form that can be combined with time-series and image analysis results.

[0148] The step of generating a single multimodal fusion vector reflecting the correlation weights between heterogeneous data by simultaneously inputting time-series latent vectors, spatial behavioral latent vectors, and text embedding vectors into a cross-attention network layer is a step that combines feature representations derived from different modalities into a single comprehensive representation by learningly reflecting the relationships between them. Here, the cross-attention network layer may refer to a neural network structure that cross-references features by calculating, in the form of weights, how closely one type of data is related to another type of data. Furthermore, the correlation weight may be a numerical value reflecting, for example, how closely an image-based danger zone entry pattern relates to the context of "need to control lifting radius" in text, or how closely an abnormal vibration pattern in equipment logs relates to the context of "need to inspect hoisting machine" in text. The multimodal fusion vector may refer to a resulting vector that expresses the complex danger situation of the entire site in a single form by combining individual features of time-series, space, and text through the aforementioned weights.

[0149] The step of generating a single multimodal fusion vector reflecting the mutual correlation weights between heterogeneous data by simultaneously inputting time-series latent vectors, spatial behavioral latent vectors, and text embedding vectors into a cross-attention network layer can be performed by, for example, inputting together a time-series latent vector reflecting strong wind and vibration anomalies, a spatial behavioral latent vector reflecting the worker's posture of tilting outward on the railing of a high-altitude work platform, and a text embedding vector reflecting context such as working at height on an exterior wall, mandatory safety harnesses, and warning of falling objects, thereby assigning higher weights to feature combinations highly associated with the risk of falling and falling objects. In this case, the multimodal fusion vector may not be a value simply combined from the three types of latent vectors, but rather a high-dimensional feature representation reflecting the semantic, temporal, and spatial correlations between them. Accordingly, even if a risk is weakly manifested in a single dataset, if it is simultaneously supported by multiple datasets, it may be judged as having a higher risk level in the subsequent classification stage.

[0150] The step of inputting a multimodal fusion vector into a gradient-boosting-based ensemble classifier to calculate the probability of occurrence of each type of serious accident at a specific point in time as a probability index in percentage units is a step of numerically predicting the probability of occurrence of each type of serious accident, such as falls, entrapment, collision, electric shock, fire, explosion, and suffocation, at a target work site using the said multimodal fusion vector. Here, a gradient-boosting-based ensemble classifier may refer to a machine learning algorithm that improves prediction accuracy by sequentially combining multiple weak prediction models and continuously compensating for the errors of previous models. Furthermore, the probability of accident occurrence may refer to a value representing the probability of the corresponding type of accident occurring within a specific time interval or under specific process conditions as a percentage or a risk probability score.

[0151] The step of inputting a multimodal fusion vector into a gradient-boosting-based ensemble classifier to calculate the probability of accident occurrence by type of serious accident at a specific point in time as a probability index in percentage units can be calculated in the form of a probability index such as 82 percent for the risk of falling and 67 percent for the risk of collision related to falling objects, for example, when a multimodal fusion vector is input that simultaneously reflects strong winds, the worker's posture tilted outward on the handrail, the possibility of not wearing a safety harness, conditions for permitting high-altitude work, and abnormal vibration of the high-altitude work platform for high-altitude work. Additionally, when the lifting process reflects the lifting radius approach behavior, abnormal crane vibration, conditions for transporting heavy objects in the work plan, and omission of control phrases, the risk of entrapment or collision may be calculated as relatively high. At this time, the device can generate prediction results in the form of multiple outputs by simultaneously calculating probability indices for each of the multiple types of accidents.

[0152] The step of generating work risk prediction information by combining the probability of an accident with risk factors, unsafe behaviors, abnormal conditions of the site environment, and signs of equipment abnormalities is a step of generating final work risk prediction information by integrating the already identified work process-specific risk factors, unsafe worker behaviors, abnormal conditions of the site environment, and signs of equipment abnormalities together with the probability index, rather than using the probability index alone.

[0153] Here, work risk prediction information refers to a structured result that comprehensively describes what risks exist for a specific point in time or a specific work process, how those risks are combined with behavioral, environmental, and equipment factors, and the actual probability of an accident. In other words, the probability of an accident forms the numerical central axis of the final risk information, while risk factors, unsafe behaviors, abnormal conditions in the site environment, and signs of equipment anomalies can provide the contextual information from which those figures were derived.

[0154] The step of generating work risk prediction information by combining the probability of an accident with risk factors, unsafe behaviors, abnormal site environmental conditions, and signs of equipment abnormalities can be performed by, for example, by grouping exterior wall high-altitude work processes, detection of non-wearing of safety harnesses, strong winds and low temperature environments, abnormal vibrations of elevated work platforms, and an 82 percent probability of a fall accident into a single structured record to generate final work risk prediction information. As another example, work risk prediction information can be generated in the form of lifting processes, worker access within the lifting radius, low light conditions at night, hoist tension imbalance, and a 74 percent probability of a collision accident. Accordingly, in subsequent steps, rather than simple probability values, one can refer to which risk factors, which unsafe behaviors, and which environmental and equipment conditions were combined to calculate a specific accident probability, and based on this, worker behavior indicators, site condition indicators, and organizational management indicators can be calculated more precisely.

[0155] And, the step of calculating worker behavior indicators, site condition indicators, and organizational management indicators based on the above work risk prediction information comprises: generating risk event records by mapping worker identification information, work process identification information, work zone identification information, equipment identification information, and accident type identification information for each of the work process-specific risk factors, unsafe worker behaviors, abnormal site environment conditions, signs of equipment abnormalities, and accident probability included in the above work risk prediction information; clustering the risk event records based on worker units, work process units, work zone units, equipment units, and time window units, respectively, to generate risk event sets by worker, risk event sets by work zone, risk event sets by work process, and risk event sets by equipment, respectively; and for each of the risk event sets by worker, calculating the frequency of non-wearing of protective equipment, frequency of unsafe movement paths, number of entry into hazardous zones, number of failure to maintain safety distances, duration of abnormal postures, average interval between repeated unsafe behaviors, simultaneous occurrence rate of multiple unsafe behaviors by the same worker, and weighted average values ​​of accident probability. A step of performing minimum-maximum normalization on the weighted average values ​​of the frequency of non-wearing of protective equipment, frequency of unsafe movement paths, number of entry into hazardous zones, number of failure to secure safety distances, duration of abnormal postures, average interval between repetitive unsafe behaviors, simultaneous occurrence rate of multiple unsafe behaviors by the same worker, and probability of accident occurrence, and calculating worker behavior indicators by applying risk type weights, work process weights, accident severity weights, and worker role weights to the normalized values;For each of the above-mentioned sets of risk events by work zone, sets of risk events by work process, and sets of risk events by equipment, a step of calculating the degree of rapid increase in hazardous gas concentration, the degree of exceeding temperature and humidity thresholds, the duration of abnormal conditions, the intensity of abnormal equipment vibration, the simultaneous occurrence rate of multiple abnormal conditions, the spread area of ​​the risk zone, the recurrence rate of abnormal equipment signs, and the degree of overlap of multiple high-risk events in the same time period; a step of performing minimum-maximum normalization on the degree of rapid increase in hazardous gas concentration, the degree of exceeding temperature and humidity thresholds, the duration of abnormal conditions, the intensity of abnormal equipment vibration, the simultaneous occurrence rate of multiple abnormal conditions, the spread area of ​​the risk zone, the recurrence rate of abnormal equipment signs, and the degree of overlap of multiple high-risk events in the same time period, and applying weights for environmental items, weights for work processes, weights for equipment importance, and weights for disaster severity to the normalized values ​​to calculate the environmental deterioration score by work zone, the condition instability score by work process, and the abnormal concentration score by equipment; and a step of calculating the site condition index by weighting the environmental deterioration score by work zone, the condition instability score by work process, and the abnormal concentration score by equipment. A step of calculating, for all of the above risk event records, the frequency of recurrence of the same type of accident, the ratio of simultaneous exposure of multiple workers, the ratio of simultaneous high risk of multiple work processes, the spread of risk between multiple work zones, the simultaneous occurrence rate of equipment anomalies and unsafe behaviors, the number of tasks requiring management intervention, and the concentration of high-risk events;The method may include the step of performing minimum-maximum normalization on the frequency of recurrence of the same type of accident, the ratio of simultaneous exposure of multiple workers, the ratio of simultaneous high risk of multiple work processes, the degree of risk diffusion between multiple work zones, the rate of simultaneous occurrence of signs of equipment abnormality and unsafe behaviors, the number of tasks requiring management intervention, and the concentration of high-risk events, and applying weights by accident type, work process importance, work zone impact, and the degree of need for management intervention to the normalized values ​​to calculate an organizational-level risk repeatability score, an organizational-level risk diffusion score, an organizational-level simultaneous management burden score, and an organizational-level complex risk management difficulty score; and the step of calculating an organizational management indicator by weighted summing the organizational-level risk repeatability score, the organizational-level risk diffusion score, the organizational-level simultaneous management burden score, and the organizational-level complex risk management difficulty score.

[0156] The step of generating risk event records is to reorganize the work process-specific risk factors, unsafe worker behaviors, abnormal conditions of the site environment, signs of equipment abnormalities, and the probability of accidents derived from the preceding multimodal analysis into standardized units capable of calculating subsequent indicators.

[0157] Here, a hazardous event record may refer to a unit data record combining relevant identification information, risk attributes, and predicted values ​​to track who, where, and in relation to which equipment or work process there is a possibility of a specific type of accident occurring based on a single hazardous situation. The worker identification information may be an employee ID, tag number, facial recognition identifier, or access terminal identifier, and the work process identification information may be a code or name used to distinguish processes such as exterior wall painting, steel lifting, confined space inspection, fire work, or electrical inspection. Additionally, the work area identification information may refer to a zone number, floor number, block number, line number, or equipment area number, and the equipment identification information may refer to information specifying individual equipment, such as a crane number, aerial work platform number, hoist number, ventilation number, or switchboard number. The accident type identification information may be a code or label indicating which type of accident is associated with a fall, entrapment, collision, electric shock, fire, explosion, or suffocation.

[0158] The step of generating a risk event record can, for example, in a case where Foreman Kim is working near an elevated work platform without wearing a safety harness during an elevated work process on the exterior wall of Zone C, strong winds persist, an abnormal vibration of the elevated work platform is detected, and the probability of a fall accident is calculated to be 80 percent, the device can generate a single risk event record by linking the worker identification information, the work process identification information, the work zone identification information, the equipment identification information, and the accident type identification information with the risk information. At this time, the risk event record may record the predicted time, time window range, accident probability value, type of unsafe behavior, type of environmental anomaly, and type of equipment anomaly. Accordingly, risk judgment results derived from different modalities can be integrated into event data in a single format suitable for calculating subsequent indicators.

[0159] The step of clustering risk event records to generate risk event sets by worker, risk event sets by work area, risk event sets by work process, and risk event sets by equipment, respectively, is a step of reclassifying the risk event records according to a standard axis suitable for the purpose of analysis to form a data structure in sets.

[0160] Here, clustering may refer to a process of grouping events attributed to the same worker into a single worker-specific set, events belonging to the same work area into a single work area-specific set, and events corresponding to the same work process and equipment into a process-specific set and an equipment-specific set, respectively. Additionally, time window clustering may refer to a process of temporally grouping events that occurred within the same 10-minute, 30-minute, or 1-hour interval to evaluate concurrency and concentration.

[0161] The step of clustering risk event records to generate risk event sets by worker, risk event sets by work zone, risk event sets by work process, and risk event sets by equipment, respectively, can, for example, group the events of failure to wear a safety harness, entry into a hazardous zone, and abnormal posture near a high-altitude work platform that occurred to Worker A during the exterior wall process in Zone 3 from 8:00 AM to 8:30 AM into a single risk event set by worker. Additionally, strong wind warnings, falling object risk predictions, lifting radius approach events, and abnormal equipment vibration events that occurred across Zone 3 during the same time period can be grouped into a risk event set by work zone. The multiple event sets generated in this way can be utilized as input data to calculate worker behavior indicators, site condition indicators, and organizational management indicators from separate perspectives.

[0162] For each set of risk events per worker, the step of calculating the weighted average values ​​of the frequency of non-wearing of protective equipment, frequency of unsafe movement paths, number of entries into hazardous areas, number of failures to maintain safety distances, duration of abnormal postures, average interval between repetitive unsafe behaviors, simultaneous occurrence rate of multiple unsafe behaviors by the same worker, and probability of accidents is a step of statistically calculating the distribution and intensity of risk events attributed to a specific worker to derive detailed factors representing the worker's level of personal safety compliance and risk exposure.

[0163] Here, the frequency of non-wearing of protective equipment may refer to the number of events in which protective equipment, such as safety helmets, safety belts, safety vests, or protective gloves, is determined to be unworn. The frequency of unsafe movement paths may refer to the number of times dangerous movement paths are detected, such as crossing prohibited areas, violating one-way traffic, approaching equipment from behind, or passing through areas with a risk of falling objects. The number of entries into hazardous areas may refer to the number of times a predefined lifting radius, area where access to rotating bodies is prohibited, hazardous area in confined spaces, or hazardous area in electrical hazard zones has been entered, and the number of failures to maintain a safe distance may refer to the number of times the distance to heavy equipment or hazardous facilities is determined to be below the permissible threshold distance. The duration of abnormal posture may refer to the total time that a posture with the waist excessively bent, a posture tilted outward from a handrail, a working posture on an unstable platform, or an excessive reach posture is maintained. The average interval between repetitive unsafe behaviors may be a value indicating how frequently the same worker repeats consecutive unsafe behaviors, and the simultaneous occurrence rate of multiple unsafe behaviors by the same worker may refer to the ratio of two or more unsafe behaviors occurring together within a single time window. Additionally, the weighted average of accident probability may refer to the average value obtained by applying importance weights based on behavior type to the accident probability assigned to each event.

[0164] For each set of risk events per worker, the step of calculating the weighted average values ​​for the frequency of failure to wear protective equipment, the frequency of unsafe movement paths, the number of times entering a hazardous area, the number of times safety distances were not maintained, the duration of abnormal postures, the average interval between repeated unsafe behaviors, the simultaneous occurrence rate of multiple unsafe behaviors by the same worker, and the probability of an accident may, for example, result in: failure to wear a safety helmet 2 times, entry into the lifting radius 3 times, failure to maintain safety distance 4 times, an accumulated abnormal posture of 90 seconds, an average interval between unsafe behaviors of 4 minutes, a simultaneous occurrence rate of multiple unsafe behaviors of 0.35, and a weighted average of the probability of an accident of 0.72 over a 30-minute period for worker A. Here, the simultaneous occurrence rate of multiple unsafe behaviors may refer to the rate at which two or more safety violations are performed at once, such as a worker walking into a hazardous area while looking at a smartphone, and the shorter the average interval between repeated unsafe behaviors, the more likely it is to be evaluated as having a habitual tendency to disregard safety.

[0165] The step of calculating worker behavior indicators by applying risk type weights, work process weights, accident severity weights, and worker role weights to normalized values ​​is a step of converting the above worker-specific risk behavior characteristic values ​​to the same scale, and then converting them into a single integrated behavior indicator by reflecting the relative importance and contextual risk of each item.

[0166] Here, minimum-maximum normalization refers to a process of scaling numerical values ​​with different units, such as frequency, time, and ratio, into a relative ratio between a pre-set minimum and maximum reference value, thereby scaling them into a range of zero to one or zero to one hundred. Risk type weights may be values ​​intended to reflect different levels of risk for each type of accident, such as fall-related actions, electric shock-related actions, or fire-related actions, while work process weights may be values ​​intended to reflect the high risk inherent in the process itself, such as working at heights, working with fire, working in confined spaces, electrical work, or lifting operations. Accident severity weights may be values ​​reflecting the expected scale of damage or severity if the corresponding action leads to an actual accident, and worker role weights may be values ​​intended to reflect differences in safety levels required for each role, such as general workers, signalmen, equipment operators, supervisors, or site managers.

[0167] The step of calculating the worker behavior indicator by applying risk type weights, work process weights, accident severity weights, and worker role weights to normalized values ​​allows the device to calculate an upward worker behavior indicator by assigning high weights to these values, for example, if the frequency of failure to wear protective equipment, the number of times entering hazardous zones, and the number of times safety distances are not maintained are all normalized to a high degree, the work corresponds to work at height or lifting operations, and the worker's role corresponds to a signalman or a worker working near equipment. Conversely, relatively low weights may be applied to minor behaviors that occur temporarily in low-risk processes. Accordingly, the worker behavior indicator can be calculated not merely as the sum of the number of behaviors, but as a value that reflects behavioral risk, process characteristics, accident severity, and role importance.

[0168] For each set of risk events by work zone, work process, and equipment, the step of calculating the degree of rapid increase in hazardous gas concentration, the degree of exceedance of temperature and humidity thresholds, the duration of abnormal conditions, the intensity of abnormal equipment vibration, the simultaneous occurrence rate of multiple abnormal conditions, the diffusion area of ​​the risk zone, the recurrence rate of equipment abnormal signs, and the overlap of multiple high-risk events within the same time period is a step of quantifying instability factors related to the physical site location, the process being performed itself, and the equipment in operation, rather than individual human factors of the worker.

[0169] Here, the degree of rapid increase in hazardous gas concentration may refer to the magnitude or rate of increase relative to the standard concentration, and the degree of exceeding temperature and humidity thresholds may refer to the amount of deviation in temperature and humidity exceeding the allowable range for each process. The duration of abnormal conditions may refer to the total time that abnormal conditions, such as strong winds, low light intensity, high temperature, insufficient ventilation, or an increase in hazardous gases, persist. The intensity of equipment vibration abnormalities may refer to the increase in amplitude, frequency distortion, or intensity of impulse vibration compared to the normal vibration baseline. The simultaneous occurrence rate of multiple abnormal conditions may refer to the ratio of two or more environmental or equipment abnormalities occurring simultaneously within the same time window, and the hazardous zone diffusion area may refer to the area or ratio of the region where hazardous gases, fire signs, or other high-risk conditions have spatially expanded from the point where the initial abnormality was detected. The equipment abnormality recurrence rate may refer to the frequency with which abnormal events repeatedly occur in the same equipment over a certain period, and the overlap of multiple high-risk events in the same time period may be a numerical value indicating how much multiple high-risk events overlap at a specific point in time, such as, for example, a gas leak alarm sounding simultaneously with the detection of spark signs in nearby equipment.

[0170] The step of calculating environmental deterioration scores by work zone, condition instability scores by work process, and abnormal concentration scores by equipment by applying weights for environmental items, work processes, equipment importance, and accident severity to normalized values ​​is a step of converting the above-mentioned site and equipment-related characteristic values ​​into a standard scale, and then assigning weights that reflect site specificity and the linkage with major accidents to convert them into intermediate evaluation scores at the space, process, and equipment levels.

[0171] Here, the weights for environmental items may be values ​​reflecting the relative impact that items such as hazardous gases, wind speed, temperature, humidity, illuminance, or ventilation conditions have on safety. The weights for work processes may be values ​​intended to reflect the fact that even under the same abnormal condition, the level of risk varies depending on process characteristics, such as working in confined spaces, working with fire, or working at heights. The equipment importance weights may be values ​​intended to assign greater importance to critical equipment, such as tower cranes or hoists, where a single abnormality could lead to a major disaster, and the disaster severity weights may be values ​​intended to reflect the scale of damage if the abnormal condition leads to an actual disaster. Accordingly, the environmental deterioration score by work zone may be a value indicating the environmental instability of a specific zone, the condition instability score by work process may be a value indicating the risk of the execution status of a specific process, and the abnormality concentration score by equipment may be a value indicating how much risk signs are concentrated in a specific piece of equipment.

[0172] The step of calculating the site condition index by weighted summing the environmental deterioration score by work zone, the condition instability score by work process, and the abnormal concentration score by equipment is a step of generating a single index that comprehensively represents the physical, process, and equipment risk status of the work site.

[0173] Here, weighted summation refers to a process of multiplying the environmental deterioration score by work zone, the condition instability score by work process, and the abnormality concentration score by equipment by a predefined ratio value, and then summing them. For example, in confined space operations, the reflection ratio of the environmental deterioration score by work zone can be set high, and in heavy equipment-centered operations, the reflection ratio of the abnormality concentration score by equipment can be set relatively high. Accordingly, the site condition index can be calculated as a comprehensive condition value that reflects not only the simple environmental score but also the instability of the work process and the concentration of equipment abnormalities.

[0174] The step of calculating the frequency of recurrence of the same type of accident, the rate of simultaneous exposure of multiple workers, the rate of simultaneous high-risk for multiple work processes, the spread of risk between multiple work zones, the simultaneous occurrence rate of equipment anomalies and unsafe behaviors, the number of tasks requiring management intervention, and the concentration of high-risk events for all risk event records is intended to evaluate the management burden and complex risk structure at the site-wide or organizational level, going beyond microscopic analysis at the individual worker or individual zone level.

[0175] Here, the frequency of repeated occurrences of the same type of accident may refer to the number of times the same type of event—such as a fall, entrapment, electric shock, or fire—occurs repeatedly. The simultaneous exposure rate of multiple workers may refer to the ratio of two or more workers being simultaneously exposed to the same or similar risks within the same time period, and the simultaneous high-risk rate of multiple work processes may refer to the ratio of two or more processes being assessed as being in a high-risk state during the same time period. The risk diffusion rate among multiple work zones may refer to the extent to which a hazardous event propagates or expands from one zone to adjacent zones or multiple zones. The simultaneous occurrence rate of equipment anomalies and unsafe behaviors may refer to the ratio of dangerous worker behaviors and equipment anomalies appearing simultaneously; this value may reflect a decline in the control of the management system or a loss of control at the site. The number of tasks requiring management intervention may refer to the number of tasks determined to require immediate action, approval review, work stoppage, or safety retraining, and the high-risk event concentration may refer to a value indicating how densely high-risk events occur within a specific time period or specific zone.

[0176] The step of calculating the organizational-level risk repeatability score, organizational-level risk diffusion score, organizational-level simultaneous management burden score, and organizational-level complex risk management difficulty score by applying weights by disaster type, work process importance, work area impact, and management intervention necessity to the normalized values ​​is a step of converting the above organizational unit characteristic values ​​into detailed scores that can be interpreted from a management perspective.

[0177] Here, the organizational-level risk repeatability score may be a value indicating the degree to which the same type of risk appears repeatedly, and the organizational-level risk diffusion score may be a value indicating the degree to which risk spreads to multiple zones and processes. The organizational-level simultaneous management burden score may be a value quantifying an overload state where immediate response becomes difficult with limited safety management personnel due to the overlap of multiple situations requiring intervention within the same timeframe. Furthermore, the organizational-level complex risk management difficulty score may refer to the overall management difficulty resulting from the combination of the diversity of accident types, the importance of the process, the spatial diffusion of risk, the simultaneity of equipment malfunctions and unsafe behaviors, and the difficulty of management intervention. For example, if work at heights and hot work are performed simultaneously, multiple workers are exposed in overlapping areas, and signs of equipment malfunction are also present, the aforementioned complex risk management difficulty score may be calculated to be high.

[0178] The step of calculating organizational management indicators by weighted summing the organizational risk repetition score, organizational risk diffusion score, organizational simultaneous management burden score, and organizational complex risk management difficulty score is a step that converts the overall operational soundness of the site and the burden level of the management system into a single indicator. Here, the organizational management indicator may be a value representing the degree of complex risk and the operational capability of the safety system across the entire site operation, rather than specific workers or equipment. The aforementioned weighted summing can be performed by applying different ratios according to the importance of each detailed score; for example, the weight of the risk repetition score may be increased in accident-prone sites, while the weight of the risk diffusion score and simultaneous management burden score may be increased in sites where multi-zone simultaneous work is frequent. Accordingly, the aforementioned organizational management indicators can be utilized as a basis for determining the priority of managerial intervention, reallocating resources, strengthening work approval policies, supplementing the safety management system, and deciding the level of response for preventing serious accidents.

[0179] An apparatus according to one embodiment includes a processor and memory. The processor may include at least one apparatus described above through the drawings or perform at least one method described above through the drawings. The memory may store information related to the method described above or store a program in which the method described above is implemented. The memory may be volatile memory or non-volatile memory.

[0180] The processor can execute a program and control the device. The code of the program executed by the processor can be stored in memory. The device can be connected to an external device (e.g., a personal computer or a network) through an input / output device (not shown in the drawing) and exchange data.

[0181] The embodiments described above may be implemented as hardware components, software components, and / or combinations of hardware and software components. For example, the devices, methods, and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are also possible.

[0182] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiment, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

[0183] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or instruct the processing unit independently or collectively. Software and / or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium, or device so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0184] Although the embodiments have been described above with reference to the limited drawings, those skilled in the art can apply various technical modifications and variations based on the above. For example, suitable results may be achieved even if the described techniques are performed in a different order than described, and / or if the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0185] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below. Explanation of the symbols

[0186] S100: Step for acquiring work-related data S200: Step for generating integrated input data for analysis S300: Step for deriving work risk prediction information S310: Step for extracting time series latent vectors S320: Step for extracting spatial behavior latent vectors S330: Step to extract text embedding vectors S340: Step to generate a multimodal fusion vector S350: Step of calculating the probability of accident occurrence by type of serious accident as a probability index in percentage units. S360: Step for generating work risk prediction information S400: Step for calculating worker behavior indicators, site condition indicators, and organizational management indicators S401: Step to generate a risk event record S402: A step of generating a risk event set by worker, a risk event set by work area, a risk event set by work process, and a risk event set by equipment, respectively. S403: A step for calculating the weighted average of the frequency of non-wearing of protective equipment, frequency of unsafe movement paths, number of entries into hazardous areas, number of failures to maintain safety distances, duration of abnormal postures, average interval between repetitive unsafe behaviors, simultaneous occurrence rate of multiple unsafe behaviors by the same worker, and the probability of accidents. S404: Step for calculating worker behavior indicators S405: A step for calculating the degree of rapid increase in hazardous gas concentration, the degree of exceedance of temperature and humidity thresholds, the duration of abnormal conditions, the intensity of abnormal equipment vibration, the simultaneous occurrence rate of multiple abnormal conditions, the diffusion area of ​​the danger zone, the recurrence rate of equipment abnormal signs, and the degree of overlap of multiple high-risk events in the same time period. S406: Step for calculating environmental deterioration scores by work zone, condition instability scores by work process, and abnormal concentration scores by equipment. S407: Step for calculating field condition indicators S408: A step for calculating the frequency of recurrence of the same type of accident, the rate of simultaneous exposure of multiple workers, the rate of simultaneous high-risk for multiple work processes, the spread of risk between multiple work zones, the rate of simultaneous occurrence of equipment anomalies and unsafe behaviors, the number of tasks requiring management intervention, and the concentration of high-risk events. S409: A step for calculating the organizational risk repeatability score, organizational risk spread score, organizational concurrent management burden score, and organizational complex risk management difficulty score. S410: Step for calculating organizational management indicators S500: A step for assigning a safety rating corresponding to the integrated safety score.

Claims

Claim 1 An integrated safety management method for AI-based pre-identification of work risks and calculation of dynamic safety scores, wherein a device includes a processor, memory, a communication module, and a non-transient storage medium, and the method is performed by the processor by executing a program stored in the non-transient storage medium, comprising: a step of acquiring a plurality of work-related data from a work site, including IoT environment sensor data, CCTV video data, work plan data, work permit data, meeting audio data, safety inspection documents, and equipment log data; a step of generating integrated input data for analysis by performing preprocessing on the work-related data, including data format unification, noise removal, missing value correction, time synchronization, and normalization; a step of deriving work risk prediction information, including risk factors by work process, worker unsafe behavior, abnormal conditions of the site environment, signs of equipment abnormality, and the possibility of accident occurrence, by performing multimodal fusion analysis on the integrated input data for analysis using a pre-trained artificial intelligence model; and a step of calculating worker behavior indicators, site condition indicators, and organizational management indicators based on the work risk prediction information. The method includes the step of calculating an integrated safety score by applying pre-set weights to the worker behavior indicators, site condition indicators, and organizational management indicators, and assigning a safety grade corresponding to the integrated safety score; wherein the step of deriving work risk prediction information comprises: deriving abnormal site environment conditions and signs of abnormal equipment conditions, and extracting a time-series latent vector by applying a prediction model based on a Long Short-Term Memory (LSTM) algorithm to time-series stream data including IoT environment sensor data and equipment log data among the integrated input data for analysis, thereby analyzing patterns of rapid increase in hazardous gas concentration, exceedance of temperature and humidity thresholds, and abnormal equipment vibration within a pre-set time window.A step of identifying unsafe worker behaviors and extracting spatial behavior potential vectors by analyzing spatiotemporal behavior patterns, including the worker's failure to wear personal protective equipment (PPE), unsafe movement paths, entry into hazardous areas, and failure to maintain a safe distance from heavy equipment, by applying a vision AI model combined with object detection and 3D skeletal estimation algorithms to CCTV video data among the integrated input data for analysis; a step of extracting risk factors for each work process, including the inherent potential risks of the day's work process, and extracting text embedding vectors by performing named entity recognition (NER) and contextual analysis using an industrial safety-specific pre-trained language model (Domain-Specific Transformer) on work plan data, work permit data, text data converted from meeting audio data, and safety inspection documents among the integrated input data for analysis; A step of simultaneously inputting the time-series latent vector, spatial behavioral latent vector, and text embedding vector into a Cross-Attention network layer to generate a single multi-modal fused vector reflecting the mutual correlation weights between heterogeneous data; a step of inputting the multi-modal fused vector into a gradient-boosting-based ensemble classifier (XGBoost) to calculate the probability of accident occurrence by type of major accident, including falls, entrapment, collision, electric shock, fire, explosion, and suffocation, at a specific point in time at a target work site as a probability index in percentage units; and a step of generating work risk prediction information by combining the accident occurrence probability with risk factors, unsafe behaviors, abnormal conditions of the site environment, and signs of equipment abnormalities.The step of calculating worker behavior indicators, site condition indicators, and organizational management indicators based on the above work risk prediction information, including: generating risk event records by mapping worker identification information, work process identification information, work zone identification information, equipment identification information, and accident type identification information for each of the work process-specific risk factors, unsafe worker behaviors, abnormal site environment conditions, signs of equipment abnormalities, and accident probability included in the above work risk prediction information; clustering the risk event records based on worker units, work process units, work zone units, equipment units, and time window units, respectively, to generate risk event sets by worker, risk event sets by work zone, risk event sets by work process, and risk event sets by equipment, respectively; and for each of the risk event sets by worker, calculating the frequency of non-wearing of protective equipment, frequency of unsafe movement paths, number of entry into hazardous zones, number of failure to secure safety distances, duration of abnormal postures, average interval between repeated unsafe behaviors, simultaneous occurrence rate of multiple unsafe behaviors by the same worker, and weighted average values ​​of accident probability. A step of performing minimum-maximum normalization on the weighted average values ​​of the frequency of non-wearing of protective equipment, frequency of unsafe movement paths, number of entry into hazardous zones, number of failure to secure safety distances, duration of abnormal postures, average interval between repetitive unsafe behaviors, simultaneous occurrence rate of multiple unsafe behaviors by the same worker, and probability of accident occurrence, and calculating worker behavior indicators by applying risk type weights, work process weights, accident severity weights, and worker role weights to the normalized values;For each of the above-mentioned sets of risk events by work zone, sets of risk events by work process, and sets of risk events by equipment, a step of calculating the degree of rapid increase in hazardous gas concentration, the degree of exceeding temperature and humidity thresholds, the duration of abnormal conditions, the intensity of abnormal equipment vibration, the simultaneous occurrence rate of multiple abnormal conditions, the spread area of ​​the risk zone, the recurrence rate of abnormal equipment signs, and the degree of overlap of multiple high-risk events in the same time period; a step of performing minimum-maximum normalization on the degree of rapid increase in hazardous gas concentration, the degree of exceeding temperature and humidity thresholds, the duration of abnormal conditions, the intensity of abnormal equipment vibration, the simultaneous occurrence rate of multiple abnormal conditions, the spread area of ​​the risk zone, the recurrence rate of abnormal equipment signs, and the degree of overlap of multiple high-risk events in the same time period, and applying weights for environmental items, weights for work processes, weights for equipment importance, and weights for disaster severity to the normalized values ​​to calculate the environmental deterioration score by work zone, the condition instability score by work process, and the abnormal concentration score by equipment; and a step of calculating the site condition index by weighting the environmental deterioration score by work zone, the condition instability score by work process, and the abnormal concentration score by equipment. A step of calculating, for all of the above risk event records, the frequency of recurrence of the same type of accident, the ratio of simultaneous exposure of multiple workers, the ratio of simultaneous high risk of multiple work processes, the spread of risk between multiple work zones, the simultaneous occurrence rate of equipment anomalies and unsafe behaviors, the number of tasks requiring management intervention, and the concentration of high-risk events;A method for AI-based pre-identification of work risks and dynamic safety score calculation, comprising the steps of: performing minimum-maximum normalization on the frequency of recurrence of the same type of accident, the ratio of simultaneous exposure of multiple workers, the ratio of simultaneous high risk of multiple work processes, the degree of risk diffusion between multiple work zones, the rate of simultaneous occurrence of equipment anomalies and unsafe behaviors, the number of tasks requiring management intervention, and the concentration of high-risk events; applying weights by accident type, work process importance, work zone impact, and the degree of need for management intervention to the normalized values ​​to calculate an organizational-level risk repeatability score, an organizational-level risk diffusion score, an organizational-level simultaneous management burden score, and an organizational-level complex risk management difficulty score; and calculating an organizational management indicator by weighted summing the organizational-level risk repeatability score, the organizational-level risk diffusion score, the organizational-level simultaneous management burden score, and the organizational-level complex risk management difficulty score. Claim 2 delete Claim 3 delete

Images