알림 목록

Analysis Area:  Ulsan Metropolitan City

Core Areas:  Mipo-Onsan National Industrial Complex, Nam-gu, Dong-gu, Buk-gu, Ulju-gun

Agenda: Levels of Prediction, Dissemination, and Protection of Labor, Safety, and Health Data

Golden Time Type: Post-accident Data Lock-in Risk

Reference Date: 2026.08.28

Version: Regional AX Golden Time Intelligence v3.2
 

In 2025, the number of fatal accidents subject to investigation nationwide was 605, an increase of 16 from the previous year, with 158 fatalities occurring in the manufacturing sector. Although the number of manufacturing fatalities decreased by 17, the number of accidents increased by 4, indicating that the scale of deaths and the frequency of occurrence followed different directions. Public data linking risks by industry, company size, prime contractors/subcontractors, and process in Ulsan in real time is not available. If safety is judged solely based on the total number of fatalities for the 2026–2028 period, warning signs of increasing risks within the process will only be identified after an accident has occurred.  The starting point of Industrial Safety AX is not the digitization of accident statistics, but rather the availability of data prior to accidents.

Ulsan secured 6.6 billion won in state funding for an AI and big data-based heatwave response project by 2026 and established an integrated control center for smart green industrial complexes. While safety management has shifted from patrols and reporting to a system based on sensors, spatial information, and predictive data, the scope of integration—which includes worker exposure, process risks, and health information at industrial sites—has not been disclosed. If urban disaster AI and workplace safety data remain separated over the next two to three years, it will be unable to predict complex risks associated with heatwaves, chemicals, heavy objects, and confined spaces. Although  Ulsan's safety AI has entered the urban environment phase, it has not yet reached the stage of integration at the workplace.

In 2025, workplaces with fewer than 50 employees accounted for 351 of the total fatal accidents, or approximately 58% of the total, while workplaces with fewer than 5 employees recorded 174 deaths, a 14.5% increase from the previous year. Although the structure is such that industrial safety risks are concentrated in small-scale sites—where management personnel and facility investments are weaker—than in large-scale workplaces, the adoption rate of AI safety equipment by company size among subcontractors in Ulsan has not been verified. If the spread of smart safety centered on large corporations continues from 2026 to 2028, risks will shift to the lower end of the supply chain rather than be eliminated.  The biggest gap in Industrial Safety AX is not the level of technology, but the disparity in access based on company size.

The Korea Occupational Safety and Health Agency (KOSHA) is providing smart safety equipment, including AI video recognition, smart crane collision avoidance, forklift proximity warnings, and collapse and displacement detection. Although safety technology has shifted from protective gear and signs to real-time detection, warnings, and automatic stops, the number of installed workplaces in Ulsan, false alarm rates, accident reduction rates, and continuous equipment operation rates are not disclosed. If the number of installations increases over two to three years without verified risk reduction effects, the equipment remains merely a purchase record rather than a measure of safety performance.  Smart safety technology has been commercialized, but its performance in the field in Ulsan remains unconfirmed.

Traditional safety management aggregated accidents, near misses, checklists, and health examination results retrospectively. While AI safety is structured to combine worker location, equipment vibration, gas concentration, temperature, work sequence, and past accidents in real time, data standards and the scope of joint analysis for the Ulsan Industrial Complex remain unconfirmed. If data for 2026–2028 remains confined to individual company equipment, the same types of accidents will be repeated at other workplaces.  The unit of competition in industrial safety has shifted from individual equipment to the risk learning speed of the entire supply chain.

Unlike accidents, occupational health issues manifest as diseases after the long-term accumulation of effects from noise, dust, hazardous chemicals, repetitive movements, and shift work. While data on health examinations and approved industrial accidents exist, regional data linking daily exposure with individual long-term health changes is not disclosed. If business restructuring and workforce mobility occur over the next two to three years, exposure history will be dispersed across multiple workplaces, making it more difficult to trace causal relationships.  The gap in Occupational Health AX lies not in diagnostic technology, but in the exposure history that persists beyond the movement of workers between workplaces.

The Korea Occupational Safety and Health Agency’s serious accident cases in the shipbuilding industry reveal an accumulation of recurring risks, such as scaffolding, heavy loads, rope breakage, falls, and collisions. Since shipbuilding production consists of blocking, welding, painting, and assembly, as well as multi-stage cooperative work, risks occur at process boundaries; however, no public disclosure system integrating the locations, work permits, and equipment status of prime contractors and subcontractors has been identified. As order volumes and the deployment of foreign labor increase over two to three years, information gaps between work shifts raise the probability of accidents.  The AX GAP in shipbuilding safety lies not in a lack of knowledge regarding risk types, but in the inability to control simultaneous operations in real time.

The Korea Occupational Safety and Health Agency analyzed exposure to hazardous factors and occupational cancer risks among welding, painting, and production workers in the shipbuilding industry. Although risk assessments have expanded beyond accidents to include long-term diseases, there is no publicly available cohort linking the cumulative exposure by job function of Ulsan shipbuilders with their post-retirement health outcomes. Even if work automation proceeds until 2028, diseases among those with past exposures will appear later.  Shipbuilding AX cannot immediately eliminate the health debt arising from past exposures, even if it reduces current accidents.

Petrochemical complexes operate flammable and toxic substances, high-pressure and high-temperature facilities, and continuous processes. Although risks arise from a combination of equipment degradation, process deviations, and maintenance schedules rather than the behavior of a single worker, the level of risk data sharing regarding piping and utilities among companies is not disclosed. If maintenance cycles and workforce compositions change due to production cuts and facility reorganization between 2026 and 2028, the predictive power of models trained on normal operation data may decrease.  The risk for Petrochemical Safety AX lies not only in the absence of a model but also in the fact that the learning environment itself changes due to restructuring.

While chemical exposure is managed through workplace environmental measurements and special health examinations, concentration fluctuations between measurement cycles and the cumulative exposure of subcontracted workers are captured only to a limited extent. Linking process sensor data with health data improves predictive power but creates gaps regarding personal information, employment discrimination, and data ownership. If analysis is expanded without rights rules for two to three years, there is a high risk that health predictions will be used for worker selection rather than for work improvement.  Occupational Health AX is technically feasible, but it is not deemed a protected technology without labor data rights.

In automobile factories, robots, conveyors, presses, forklifts, and workers operate within the same production space. While smart safety equipment detects risks of approach, collision, and entrapment in real time, it is unclear whether the equipment of suppliers and logistics companies is connected to the same level. If only the interior of the finished vehicle factory is advanced over the next two to three years, the risks shift to parts loading/unloading, outsourced maintenance, and logistics sections.  The level of completeness of automotive safety AX is lower at the boundaries of the supply chain than within the factory.

The expansion of electrification and battery production adds high voltage, thermal runaway, chemical substances, and new maintenance risks. While the industrial structure is shifting from mechanical risks associated with internal combustion engines to complex risks involving electricity, chemicals, and software, it is not disclosed how much the accident and near-miss classification system at the Ulsan site reflects this shift. Learning only from past accident types until 2028 will fail to predict low-frequency, high-risk accidents in new processes.  Future Vehicle Safety AX is strong against historical risks with abundant data but vulnerable to new risks with limited data.

Ulsan has launched a project to collect and predict localized heatwaves using AI and big data. While the potential for heatwave information to be subdivided from average urban temperatures into workplace-specific microclimates has increased, the extent to which it links to workers' work intensity, protective equipment, shift times, and underlying conditions remains unconfirmed. If climate risks increase over two to three years, disparities in health damage across different processes widen even at the same temperature.  Ulsan's Heatwave AX is currently at the environmental prediction stage and not the worker-specific risk assessment stage.

The shipbuilding industry is expanding its foreign workforce, while the manufacturing sector as a whole is experiencing an aging of its skilled workforce. Although the languages, experiences, and health conditions required to understand risk information have diversified, real-time multilingual alerts and the application rates of individual risk levels are not being disclosed. If changes in the workforce structure between 2026 and 2028 outpace safety training and alert systems, warnings will not translate into action even if technology is installed.  Compared to changes in the workforce composition, the personalization and multilingualization of safety information in Ulsan remain unconfirmed.

AI video recognition and location tracking detect entry into hazardous areas and the failure to wear protective equipment. While this enhances the visibility of safety management, there are no common regional standards regarding whether workers' movement, speed, and behavior data are reused for productivity evaluations or disciplinary actions. If safety data is used for purposes other than intended for two to three years, workers choose to avoid sensors and fail to report, leading to a decline in data quality.  Safety AI, where surveillance and protection are not separated, undermines predictive performance due to a lack of trust.

While AI models learn from past accident and normal operation data, industrial accidents occur infrequently and involve complex situation combinations. Due to the scarcity of accident data, verifying false alarms and missed detections is difficult, yet performance indicators for each model at the Ulsan site are not disclosed. By 2028, if alarm accuracy is low, the site will ignore repeated alarms and disable the automatic stop function.  The evaluation criteria for safety AI are not installation status, but rather the missed detection rate and site reliability.

Ulsan possesses the Smart Green Industrial Complex Integrated Control Center, AI heatwave forecasting, a dedicated industrial safety organization, and smart safety systems from large corporations. While readiness at the level of the city, industrial complexes, and large enterprises is confirmed, the levels of equipment, data, and dedicated personnel at partner companies and small-scale workplaces are not disclosed. If readiness at the bottom of the supply chain is not measured for two to three years, the regional average is overestimated by the performance of large corporations.  Ulsan's Safety AX readiness level is judged as 'advanced' for large enterprises and 'uneven' for the industrial ecosystem.

In 2025, approximately 58% of nationwide fatal accidents occurred in workplaces with fewer than 50 employees. While the center of risk lies in small-scale workplaces, operating AI safety equipment requires investment costs, communication, maintenance, and data analysis capabilities. The sustained operation rate of small businesses in Ulsan has not been verified. If only initial subsidies are available between 2026 and 2028, the safety gap will widen again when equipment replacement and subscription costs arise.  The preparedness level of small-scale workplaces is assessed as being weaker in sustained operation than in installation.

AI video recognition, smart cranes, forklift approach detection, and collapse/displacement sensors have been included in the commercial support items. While various types of technology have spread, installation rates and accident reduction effects by industry and process in Ulsan are not disclosed. If the distribution of individual equipment continues for only two to three years, process accidents involving the combination of multiple risks cannot be predicted.  Smart safety equipment has spread to the product market, but it is currently in a state of partial diffusion across the entire Ulsan process.

The Smart Green Industrial Complex Integrated Control Center established the foundation for linking 3D spatial information, control, and services by 2025. While visibility at the industrial complex level has improved, real-time linkage with internal work permits, equipment sensors, and worker risk information has not been confirmed. If external control and internal process data are separated by 2028, location identification after an accident will be faster, but intervention prior to an accident will be limited.  Industrial complex control has expanded, but the connection of predictive safety data remains incomplete.

Large corporations possess their own safety organizations, sensors, control rooms, and the capability to invest in facilities. Even when subcontractors work at the prime contractor's site, data by affiliation, process, and equipment may be separated, and the rate of shared data operation is not disclosed. If safety data between the prime contractor and subcontractors is not linked for two to three years, both responsibility and risk regarding process boundaries are simultaneously omitted.  The prime contractor's safety AX has been implemented, but the joint prime contractor-subcontractor AX remains unconfirmed.

The proportion of fatal accidents in workplaces with fewer than 50 employees has exceeded half. Although high-risk groups do not coincide with groups capable of technology investment, Ulsan's performance in support, installation, and maintenance by company size has not been verified. If this gap persists until 2028, the overall regional accident risk will remain at the bottom of the supply chain despite advancements in safety technology.  The level of diffusion among partner companies is assessed to be significantly lower than that of large corporations.

Occupational health manages hazardous factors and diseases through workplace environment measurements, special health examinations, and epidemiological investigations. While institutional data has been accumulated, a Ulsan-level system linking individual exposure, workplace mobility, and health outcomes over the long term has not been identified. If the turnover of jobs, retirements, and foreign workers increases over a period of two to three years, the discontinuity in exposure history becomes even greater.  Health checkups have become widespread, yet lifetime cumulative exposure data remains fragmented.

Research on the shipbuilding industry analyzed the risk of occupational diseases by linking job-exposure matrices with data from cancer registration, employment insurance, and health examinations. While the research methodology exists, the extent to which it is fed back to on-site workers in the form of individual risk alerts and work adjustments has not been verified. If research and the workplace become separated by 2028, disease risks will remain solely as collective statistics.  Although the capacity for analyzing occupational health exists, the spread of real-time prevention is still in a pre-initial stage.

From 2026 to 2028, Ulsan will simultaneously experience an increase in shipbuilding orders, automotive electrification, petrochemical restructuring, climate risks, and an expansion of the foreign workforce. While workloads, processes, facilities, and workforce composition will all change together, the update cycle for historical data-based models to reflect these new conditions is not disclosed. If change outpaces learning, the accuracy of safety AI declines during the most critical transitional period.  The time-sensitive risk for Ulsan Safety AX lies not in the delay of technology adoption, but in the delay of model updates regarding changes in the industrial structure.

Occupational diseases involve a time lag of several years between exposure and diagnosis. If facilities and employment structures change over the next two to three years, current workers' exposure information may become fragmented due to workplace closures and job changes. Consequently, it becomes difficult to identify the cause and assign responsibility even if a disease occurs later.  The current period, amidst the restructuring of the petrochemical and shipbuilding industries, is the final opportunity to preserve occupational health histories.

Serious accidents operate on a structure where data accumulates only after the loss of life and health has occurred. Even if models are improved after an accident, losses already incurred cannot be restored, and on-site memory weakens due to the replacement of workers in the same process. If near-miss incidents and warning data are not accumulated for two to three years, the AI ​​continues to use data on deaths and injuries as training material.  The irreversibility of industrial safety lies in the fact that irrecoverable costs are incurred the moment failure data is generated.

If health data is fragmented by company, it becomes difficult to recover past exposures after a worker changes jobs or retires. As business restructuring and the closure of partner companies proceed, records of the working environment and the responsible parties may disappear along with them. If long-term histories are not linked by 2028, the likelihood of determining causality for occupational diseases will be permanently reduced.  The discontinuity of labor health data is an irreversible risk that simultaneously undermines health damages and the right to compensation.

Compression judgment: For Ulsan’s Labor, Safety, and Health AX to be established, not equipment supply, but near-miss detection rate, alarm lead time, partner connection rate, cumulative exposure tracking rate, and accident reduction must be verified simultaneously before 2028.

It is a stage, and the predictive system linking workers' lifetime exposure, prime contractor/subcontractor processes, and health outcomes is judged to be incomplete.

While smart safety is being implemented at large corporations, there is a lack of evidence that it has spread to the same level to subcontractors, small workplaces, and foreign workers where risks are concentrated.

Labor AX must be a system that determines risks prior to accidents and exposure prior to diseases earlier, rather than a system that monitors workers more precisely. Currently, Ulsan possesses detection technology but has not reached the simultaneous verification of prediction, diffusion, and rights.

Golden Time Judgment: Smart Safety Adoption + Predictive Labor-health Integration Risk

Evidence Sources

• Ministry of Employment and Labor, Status of Fatal Accidents Subject to Accident Investigation in 2025
• Ministry of Employment and Labor, 2025 Industrial Accident Status
• Korea Occupational Safety and Health Agency, Serious Accident Statistics
• Korea Occupational Safety and Health Agency, serious accident cases in the shipbuilding industry
• Korea Occupational Safety and Health Agency analyzes hazardous factors and occupational diseases in the shipbuilding industry
• Korea Occupational Safety and Health Agency supports smart safety equipment
• Korea Occupational Safety and Health Agency, Field Cases of AI and Smart Industrial Safety Technology
• Ulsan Metropolitan City, AI and Big Data-based Heatwave Response
• Ulsan Metropolitan City, Smart Green Industrial Complex Integrated Control Center

Structural Insight — Industrial Safety AI is Already Too Late the More It Learns from Accidents

The performance of general manufacturing AI improves as the amount of data on production volume, defects, and maintenance increases. While industrial safety AI models can become more sophisticated as the number of accidents to learn from grows, that data is generated at the cost of injuries and fatalities. It is a unique field where data accumulation and social outcomes do not move in the same direction.

The rarer accidents are, the more difficult statistical learning becomes, but the higher the correlation with safety. Conversely, accident-prone sites are favorable for model training but represent locations where prevention has already failed. Therefore, if industrial safety AI is designed around accident data, a paradox arises where the safest factories lack sufficient training data, while the most dangerous factories show improved predictive power only after damage has accumulated.

The data that changes this structure is not deaths or injuries, but rather equipment deviations, work overlap, ignored warnings, detachment of protective equipment, near misses, and exposure to hazardous factors. However, this data is scattered within the company and overlaps with worker surveillance information. While the preventive value increases the earlier it is collected before an accident, the risk of rights infringement also grows.

Therefore, the structural turning point of Ulsan Labor AX is not the accuracy of the AI ​​itself.  Whether Industrial Safety AX can be established depends on whether data to learn risks without human injury and the trust that such data is used solely for worker protection are simultaneously secured.