Intelligent chemical production mode occupational health risk identification, assessment and intervention study
Figures
Abstract
The intelligent transformation of the chemical industry has reshaped petrochemical production workflows and altered occupational-health risk landscapes. This paper delineates four typical operating scenarios for intelligent chemical manufacturing and develops a four-tier hierarchical risk indicator system covering objectives, criteria, factors, and measurable indicators. A hybrid entropy-weighted Bayesian-network evaluation model is established, where entropy weights calibrate prior probabilities of root-node inputs while conditional-probability tables are retained to characterize risk-coupling behaviors. Model performance is validated via time-series partitioning, multi-model benchmark comparisons, and cross-validation using Monte-Carlo simulations. An entropy-weighted TOPSIS multi-criteria decision-making method is adopted to rank intervention-target priorities, and discrete time-series analysis quantifies outcomes from a six-month closed-loop intervention program. Results demonstrate that human-machine interaction, psychological, and managerial risks dominate occupational-health hazards within intelligent chemical-production contexts. Three highly sensitive critical drivers are identified: operating-instruction error rate, risk-monitoring gap duration, and information-overload frequency. Multi-criteria decision analysis further reveals that high-sensitivity risk factors do not inherently represent optimal intervention targets. The proposed model achieves 88.2% prediction accuracy on time-series test datasets, and statistically significant overall-risk reduction is observed post-intervention. This work offers theoretical support for risk identification, equipment retrofitting, and worker-health protection for smart petrochemical sites, facilitating the synergistic advancement of chemical-industry intelligent modernization, operational safety, and employee occupational-health outcomes.
Citation: Shen YF, Xu JQ (2026) Intelligent chemical production mode occupational health risk identification, assessment and intervention study. PLoS One 21(11): e0363258. https://doi.org/10.1371/journal.pone.0363258
Editor: Fatemeh Afsharnia, Ferdowsi University of Mashhad, IRAN, ISLAMIC REPUBLIC OF
Received: July 31, 2026; Accepted: November 3, 2026; Published: November 20, 2026
Copyright: © 2026 Shen, Xu. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All data generated or analyzed in this study are included in the manuscript and Supporting Information files.
Funding: This work was supported by the Ningbo Higher Education Ideological and Political Education Research Association Key Project (Grant No. SGXSZA2502); and the Ministry of Education of China Humanities and Social Sciences Research Project (Special Project for University Counselors, Grant No. 23JDSZ3056).
Competing interests: The authors have declared that no competing interests exist.
1 Introduction
The chemical industry constitutes a cornerstone of national economic development. Its production workflows are continuous and intricate, with process media frequently exhibiting flammable, explosive, and highly toxic properties. Accordingly, occupational-health protection has long served as an indispensable core component of chemical-sector safety systems [1–3]. Enabled by Industry 4.0, human-cyber-physical systems (HCPS), and the Operator 5.0 paradigm, digital-twin platforms, smart sensors, and remote autonomous-control technologies have broadly transformed process-industry operating patterns. Traditional physically intensive on-site labor is progressively superseded by new work modalities: central-control-room remote monitoring, human-machine collaborative decision-making, and alert-driven emergency intervention [2]. Although technological advances mitigate physical-exposure hazards in field environments, risk origins have shifted from tangible physicochemical threats toward insidious, long-term human-factor-driven harms stemming from human-machine interfaces, cognitive overload, and chronic psychological stress [3–5].
Nevertheless, process-industry safety research exhibits a persistent thematic bias. The overwhelming majority of existing studies center on process safety, equipment-failure prevention, and mitigation of leakage-explosion incidents, treating human factors primarily as accident-causal variables. Relatively few investigations perform quantitative modeling and assessment that prioritize workers’ long-term occupational-health outcomes as core research objectives [6,7]. Much existing industrial-human-factors literature addresses how operator errors trigger manufacturing accidents yet rarely evaluates how intelligent-production workflows themselves chronically degrade operators’ physiological and psychological well-being [8]. This research gap is especially pronounced within chemical process industries. While discrete-manufacturing Operator 5.0-focused studies have explored cognitive load, wearable physiological monitoring, and human-machine symbiosis for workforce wellness, high-risk petrochemical continuous-process settings still lack robust quantitative occupational-health assessment frameworks that account for both process-production constraints and real-world plant-site conditions [9–11].
To address these theoretical limitations and practical engineering bottlenecks, the present study selects an intelligent petrochemical manufacturing base as its empirical case. Drawing on document review, on-site stakeholder workshops, semi-structured professional interviews, and multi-source field-monitoring datasets, this paper constructs a four-tier progressive occupational-health-risk indicator system. A hybrid dynamic entropy-weighted Bayesian-network assessment model is built to unpack multi-factor risk-coupling evolutionary mechanisms. A multi-criteria intervention-prioritization framework is formulated, balancing risk sensitivity, implementation cost, and practical operability; a six-month closed-loop intervention experiment is implemented for empirical validation. This research fills methodological gaps for process-industry intelligent human-factor-health assessment while complying strictly with human-subject-research ethical standards to preserve participant rights and data confidentiality. Academically, this work deconstructs four distinct intelligent-manufacturing operational scenarios, analyzes coupling mechanisms between conventional and emerging hazards, and delivers a causally interpretable, data-driven hybrid-assessment model, expanding occupational-health-risk theoretical frameworks under HCPS and Operator 5.0. Practically, it supplies actionable risk-identification instruments, assessment models, and tiered closed-loop intervention strategies usable by petrochemical enterprises and regulatory authorities, advancing concurrent progress for chemical-sector intelligent upgrading, production safety, and worker occupational health. The overall technical workflow of this study is shown in Fig 1.
The research ethics statement for this paper can be found in the Supplementary Materials Part 1 and S1 Fig.
2 Scenario deconstruction and risk mechanism analysis
2.1 Classification and characteristics of smart chemical production scenarios
Table 1 compares risk-scenario characteristics between conventional and intelligent chemical production, covering production modes, risk categories, risk attributes, and risk-management challenges. Driven by digital-control technologies, automated hardware, and intelligent-interaction systems, intelligent chemical production systematically reorganizes process workflows, operational practices, and management paradigms inherited from conventional chemical manufacturing [14]. Field-based manual operations have evolved into a hybrid operational model integrating remote monitoring, intelligent decision-making, and autonomous execution. Based on equipment-automation maturity, personnel-participation patterns, work-behavior traits, and occupational-health-risk exposure modes, four core operational scenarios for intelligent chemical production are defined: intelligent sensing and monitoring, intelligent control and optimization, autonomous execution with natural-language interaction, and autonomous execution for predictive maintenance (Fig 2). Each scenario exhibits distinct operational features, human-machine-interaction patterns, and occupational-health impact pathways, serving as fundamental carriers for the generation and evolution of occupational-health risks [15].
Fig 2 categorizes four types of intelligent chemical production scenarios based on business functions. These four scenarios represent functional classifications and do not imply any chronological order, sequential execution, or physical mutual exclusion; multiple scenarios may coexist within the same production unit. The figure shows only the typical equipment, software, and human-machine interaction components corresponding to each scenario. For a complete causal evolution model of differentiated occupational health risks across the four scenarios, see S2 Fig.
These four intelligent-chemical-production scenarios are classified according to business functions and task logic, rather than physical plant geography. Accordingly, they are not mutually exclusive; multiple scenario subsystems can co-exist within a single intelligent chemical facility [16]. For instance, within a chemical complex, plant-process control corresponds to Scenario 2 (Intelligent Control and Optimization); natural-language or gesture-driven operator interventions for autonomous workflows inside the central-control room correspond to Scenario 3 (Autonomous Execution and Natural Interaction); on-line equipment-status monitoring falls under Scenario 1 (Intelligent Sensing and Monitoring); and equipment-fault prediction plus maintenance alerts belong to Scenario 4 (Autonomous Execution and Predictive Maintenance). Scenario 2 emphasizes global process-parameter optimization dominated by human-initiated decision-making. In contrast, Scenario 3 focuses on emergency response for pre-automatized workflows, where personnel primarily react passively toward abnormal events [17–19]. To further clarify conceptual boundaries, Table 1 defines and contrasts the four scenarios from the perspectives of business objectives, equipment architecture, human-machine-interaction modes, typical job positions, and initial risk triggers.
S2 Fig presents causal-transmission flowcharts for all four scenarios. Each flowchart establishes a chained causal logic: intelligent-system characteristics → shifts in human-machine behaviors → intermediate risk states → final occupational-health hazards. S1 Table captures the essential distinction between Scenario 2 and Scenario 3 rooted in business objectives and personnel authority. In Scenario 2, personnel actively launch process-optimization actions; in Scenario 3, personnel perform no global-scale process optimization and only conduct anomaly-oriented emergency interventions for pre-autonomous systems. Even when both scenarios run on the same physical central-control workstation, their underlying business logic and risk triggers remain entirely separate [20]. This scenario-driven classification directly aligns with enterprise-level management requirements for diverse intelligent-production workflows. It delivers targeted support for scenario-tailored occupational-health interventions, including workstation redesign for monitoring-oriented scenarios, human-machine safety-distance specifications for collaborative scenarios, and psychological-counseling provisions for autonomous-operation scenarios [21].
2.2 Mechanisms underlying occupational health risks in smart scenarios
Occupational-health risks within intelligent chemical-production environments do not represent simple superposition of conventional hazards. Instead, they constitute a gradual evolutionary process shaped by technological upgrades, operational-model transformations, and system-level defects. The core mechanism originates from human-machine-relationship reconstruction enabled by intelligent modernization, which reshapes job-activity patterns and subsequently induces emerging health hazards. Meanwhile, coupling between intelligent-system vulnerabilities and pre-existing conventional hazards amplifies risk magnitude, expands risk scope, and generates unique risk-generation pathways specific to intelligent-manufacturing contexts.
Advancements in production-equipment intelligence constitute the fundamental driver behind transformed occupational-health-risk profiles. Automated machinery replaces manual on-site operations, substantially lowering worker exposure to chemical poisons, physical hazards, and mechanical injuries and mitigating many traditional occupational-health risks [17]. Nevertheless, operator workflows shift from physical field labor toward terminal-based monitoring and decision-making tasks, profoundly altering workers’ physiological and psychological conditions. Prolonged sedentary work yields chronic musculoskeletal strain across the neck, shoulders, lumbar, and back regions. Extended screen-watching triggers visual fatigue, ocular dryness, and degraded visual acuity. Non-standard work-rest cycles created by continuous monitoring and standby alert duties disrupt circadian rhythms and promote sleep disturbances alongside metabolic disorders. These physiological hazards stem fundamentally from revised human-machine-interaction paradigms and emerge as dominant health risks in intelligent-production settings [22].
Evolved human-machine-interaction patterns further trigger novel psychological and cognitive hazards. In intelligent-production environments, operators sustain long-duration states of information reception and decision-making execution. High-frequency alarm notifications, massive process datasets, and stringent rapid-response requirements impose persistent cognitive load. Excessive cognitive burden yields inattentiveness, slowed reaction times, and deteriorated judgment, ultimately producing cognitive fatigue [23]. Furthermore, physical separation between operators and production sites diminishes situational awareness of process conditions and increases reliance on intelligent systems. System anomalies readily provoke psychological anxiety and stress. Cumulative psychological pressure may induce chronic mental-health conditions including anxiety and depression and degrade overall workforce well-being.
Intelligent-system vulnerabilities and algorithmic biases represent critical sources of secondary occupational-health hazards. Intelligent chemical-production systems operate on sensor readings, algorithmic models, and communication networks. Sensor faults generate data distortion; algorithmic deviations produce erroneous decisions; communication latency impairs instruction-delivery timeliness [24]. These system defects not only threaten process safety but also exert direct impacts on worker occupational health. False alarms trigger repeated stress responses among personnel; missing anomaly signals leave operators unknowingly exposed to high-risk conditions. Secondary health hazards originating from system defects feature sudden onset and covert characteristics, making pre-emptive prediction and prevention difficult and increasing the overall complexity of occupational-risk governance.
Mutual interaction and amplification between legacy chemical-industry hazards and newly-emerging intelligent-context risks define the core risk-formation feature for smart-manufacturing scenarios. Legacy hazards such as toxic-gas leakage, noise pollution, and high-temperature radiation persist rather than disappearing entirely under intelligent operating conditions [25]. When combined with emerging risks such as physical strain, mental fatigue, and cognitive overload, these legacy hazards produce risk-amplification effects. System-failure-driven toxic-release events coupled with operator cognitive-fatigue-induced delayed responses magnify exposure severity and worsen health consequences. Co-occurrence of on-site inspection noise and smart-equipment-derived psychological stress accelerates hearing impairment and mental-health disorders [26]. Superimposed multi-source hazards invalidate conventional single-risk management logics developed for traditional scenarios and raise performance expectations for occupational-health risk identification and targeted intervention.
2.3 Multidimensional coupling relationships of health risks
Occupational-health risks within intelligent chemical production form a multidimensional complex system. According to their origins and action patterns, these risks fall into four major categories: physiological, psychological, environmental, and managerial risks. Extensive complex interactions exist among these categories, including pairwise coupling, multidimensional superposition, and cascading transmission. Such coupling relationships govern risk occurrence patterns, evolutionary trends, and harm severity, and constitute the core foundation for precise risk identification and targeted intervention (Fig 3). By calculating the two-factor coupling-amplification coefficient via Bayesian-network modeling (Supporting Information Section 3.3), researchers can quantitatively characterize synergistic amplification magnitudes between risk factors and generate quantitative evidence for interventions targeting coupled hazards [13,19].
Coupling between physiological and psychological risks represents the most direct and prevalent interaction mode. Physiological hazards such as musculoskeletal strain, eye fatigue, and sleep disorders commonly arise from prolonged sedentary work and irregular shift schedules. Resulting physical discomfort impairs cognitive functioning and emotional regulation, further triggering irritability, anxiety, and other psychological symptoms. Conversely, sustained psychological stress and cognitive fatigue degrade the body’s normal physiological regulatory capacity, worsening physical exhaustion and reducing overall physical fitness [20]. These two risk categories reinforce one another and create a positive-feedback loop. Minor localized discomfort may escalate into systemic chronic health conditions, substantially increasing the difficulty of risk mitigation [22].
Coupling between environmental risks and physiological risks embodies interactions between external production surroundings and human physical conditions. Environmental risks in smart manufacturing settings include residual toxic-hazardous substances, noise exposure, and temperature-humidity imbalance. Long-term exposure to adverse surroundings can directly damage worker physiological functions; for instance, toxic agents cause respiratory injury while noise induces hearing impairment. Meanwhile, physiological fatigue induced by intelligent-production workflows lowers workers’ tolerance toward environmental hazards, so that even low-concentration exposures may produce substantial health damage [23]. This coupling exhibits both immediate and cumulative properties, which readily trigger acute occupational-health injuries as well as chronic occupational diseases.
Coupling between psychological risks and managerial risks is driven by institutional mechanisms and shapes workforce mental states. Managerial risks mainly stem from inadequate early-warning frameworks, unreasonable work scheduling, incomplete training protocols, and ambiguous responsibility allocation. Unreasonable schedules plus excessive performance pressure directly raise employees’ psychological burden, whereas deficient health-management systems neglect workforce mental-health protection and allow psychological risks to accumulate continuously [24]. In turn, psychological distress among employees reduces work motivation, elevates operational-error frequencies, and degrades management effectiveness, which further amplifies managerial risks. This bidirectional interaction creates a hidden risk-transmission pathway that threatens stable operation of the entire intelligent-production system.
3. Intelligent occupational health risk identification system for smart chemical production
3.1 Overview
Occupational-health-risk identification serves as a prerequisite for risk assessment and intervention, and acts as a critical link for safeguarding personnel safety and health within intelligent chemical-production sites. Leveraging digital-twin technology, industrial internet infrastructure, intelligent sensors, automatic-control modules, and robotic operations, intelligent chemical production reshapes manufacturing workflows and operational modes. Conventional occupational-health-risk identification methodologies cannot adapt to risk-evolution characteristics under smart scenarios, suffering drawbacks such as single-faceted identification dimensions, insufficient dynamic responsiveness, and omission of coupled-risk pathways.
Intelligent chemical production integrates physical equipment, digital systems, human operations, and environmental elements; occupational-health risks accordingly present new features including multi-source origins, dynamic variation, coupling effects, and latent hazards. Risk-identification workflows must break static, one-dimensional traditional constraints. Following rigorous scientific design principles, researchers should construct an identification framework that covers full scenarios, complete workflows, and all contributing elements.
3.2 Development of an occupational health risk indicator system
Risk-indicator systems quantitatively visualize risk-identification outputs and provide the core analytical framework for risk assessment. Based on intelligent-chemical-production traits, national occupational-health standards, and risk-management requirements, this section constructs a four-tier hierarchical risk-indicator system [25]. Scientific screening and validation procedures guarantee indicator representativeness, quantifiability, and practical applicability.
Basis for constructing the indicator system.
Three core underpinnings support indicator-system development. First, relevant chemical-sector occupational-health standards and specifications ensure indicators comply with statutory monitoring and risk-management requirements for industrial chemical sites. Second, technical characteristics of intelligent chemical production match operational patterns including intelligent control, human-machine collaboration, and digital governance, highlighting the unique attributes of emerging-risk indicators. Third, core outputs from risk-identification work draw on mixed-method-derived risk factors, ensuring close alignment between indicators and real-world on-site hazards [24–26].
Furthermore, the framework follows principles of independence, systematic completeness, and practical operability. Indicator redundancy and overlap are eliminated; hierarchical logics remain clear; every indicator supports quantification via field monitoring and data collection.
Structuring the four-tier indicator system.
Fig 4a illustrates the theoretical architecture of a hierarchical Bayesian network for occupational-health-risk assessment under multi-factor coupling. This framework accepts multidimensional risk indicators as bottom-layer inputs, executes probabilistic inference plus coupling-oriented post-processing, and outputs quantified occupational-risk levels. Its key advantage lies in native capture of nonlinear cross-dimensional factor-coupling effects, avoiding subjective bias introduced by manually predefined interaction terms in classic risk models.
The bottom-level module implements a multivariate risk-indicator system that decomposes hazard sources across five dimensions: physiological risks, psychological risks, human-machine-interaction risks, secondary environmental risks, and managerial risks. The physiological dimension characterizes occupational exposure and physical fatigue; the psychological dimension quantifies anxiety severity; the human-machine-interaction dimension describes information overload and operational errors; the secondary-environmental-risk dimension reflects environmental-monitoring deficiencies; the managerial dimension captures institutional shortcomings [27]. Each indicator functions as a parent node within the Bayesian network, enabling structured input of multimodal risk information and covering human, cyber, environmental, and managerial contributing factors.
The intermediate layer adopts multi-parent-node conditional-probability tables and performs probabilistic inference based on . This architecture supports joint dependencies among multiple parent nodes without explicit manual definition of interaction terms. Cross-dimensional coupling effects are intrinsically modeled at the probability-distribution level, which differentiates it from conventional linear-superposition models. Such design accommodates complex multi-factor-interaction features of occupational-health risks and generates preliminary comprehensive occupational-health-risk ratings [28].
The coupling-effect post-processing module on the right side introduces the coupling coefficient for quantitative pairwise-coupling-strength analysis. Factor-interaction modes are categorized according to thresholds:
- : positive coupling (risk amplification)
- : independent linear superposition
- : mutual risk inhibition
A four-tier progressive indicator framework consisting of the objective layer, criterion layer, factor layer, and indicator layer is established to realize well-structured and comprehensive risk characterization (Fig 4b).
The objective layer constitutes the top hierarchical tier of the whole system, representing the overall occupational-health-risk magnitude within intelligent chemical-production environments. It intuitively quantifies comprehensive occupational-health threats to personnel under smart-manufacturing scenarios and serves as the primary target for risk assessment and mitigation.
The criteria layer forms the core dimension of the indicator system. Synthesizing risk-identification outputs and smart-production operational characteristics, it is divided into five major categories: physiological risks, psychological risks, human-machine-interaction risks, secondary environmental risks, and managerial risks. Physiological risks capture direct harm inflicted by smart operations on worker physical functions, covering both conventional physicochemical hazards and physical strain injuries induced by intelligent workflows. Psychological risks address mental stress and emotional disturbances stemming from remote operations, information surveillance, and high-pressure job demands. Human-machine-interaction risks characterize latent hazards including operational errors, command conflicts, and response delays arising from collaborative work between operators and intelligent equipment. Secondary environmental risks denote health threats triggered by abnormal environmental parameters generated by smart-device operation and production-parameter adjustment. Managerial risks reflect institutional vulnerabilities concerning intelligent-control workflows, occupational-health administration frameworks, and emergency-response mechanisms [29].
The factor layer provides granular decomposition for each criterion-layer category, with every criterion mapped to multiple core risk factors. Physiological risks are decomposed into chemical-toxin exposure, noise-vibration hazards, high-temperature-humidity impacts, and cumulative physiological fatigue. Psychological risks cover information-overload-induced stress, operational-responsibility anxiety, and monotony-driven job burnout. Human-machine-interaction risks include operational-command errors, equipment-response latency, and human-machine-interface incompatibility. Secondary environmental risks comprise smart-device thermal radiation, ventilation-system malfunctions, and uncontrolled pollutant dispersion. Managerial risks consist of insufficient risk-monitoring coverage, inadequate training programs, and incomplete emergency-response protocols [11,30].
The indicator layer represents the bottom tier of the system and contains directly quantifiable, monitor-ready specific indicators. Using the Delphi method, correlation analysis, alongside reliability-validity validation, valid quantitative indicators are screened from factor-layer entries, while hard-to-acquire and highly-correlated indicators are eliminated. This procedure yields a scientifically rigorous, practical indicator set that supports risk-factor measurement, evaluation, and governance.
The full indicator-screening workflow proceeds across four sequential rounds:
- Preliminary screening: Thirty-two candidate indicators spanning five risk dimensions are compiled from peer-reviewed literature, industrial specifications, and field investigations.
- Correlation-based elimination: Pearson correlation coefficients are computed. When , indicators with lower information contribution are discarded to mitigate multicollinearity among input variables.
- Delphi expert screening: Seven domain experts rate indicator necessity on a 1-to-5 scale. Only indicators with an average score ≥ 3.5 are retained for subsequent analysis.
- Reliability-validity testing: Indicators satisfying are preserved. Factor analysis is further implemented to verify convergent validity. Fifteen quantitative indicators are finally selected for the assessment system.
A substitute-indicator perturbation experiment is conducted, and the relative deviation of risk outputs is defined as:
(1)where
- denotes the baseline risk score calculated from the original indicator set;
- represents the risk score obtained after replacing one original indicator with an alternative candidate indicator;
- is the relative deviation (%) of risk output.
Perturbation test results reveal that 10 out of 15 indicators produce relative deviation , accounting for 66.7% of total indicators. Three extremely high-sensitivity indicators (operational-command-execution error rate, risk-monitoring-gap duration, and information-overload frequency) yield relative deviations exceeding 12%, with a maximum value reaching 14.25%. The selection of high-weight, high-sensitivity indicators exerts considerable influence on overall risk-assessment outputs [15]. Under field-monitoring-constrained conditions, low-weight indicators can be substituted with alternative indicators from the same dimension, introducing only minor disturbances to final evaluation outcomes (S2 Table).
4 A dynamic quantitative assessment model for occupational health risks
Occupational-health risks under intelligent chemical-production scenarios feature dynamic evolution, stochastic fluctuation, and multifactorial coupling. Conventional static assessment approaches fail to precisely quantify such risks or perform real-time risk simulation. Rooted in risk-dynamics and coupling principles, this chapter compares practical limitations exhibited by mainstream assessment methodologies and constructs a hybrid model integrating the entropy-weighted method and Bayesian networks [31]. This work completes model architecture establishment, parameter calibration, and scientific validation; identifies key risk drivers via sensitivity analysis; and develops a dynamic quantitative assessment framework tailored for intelligent chemical-production contexts, delivering quantitative support for risk governance and intervention practices.
4.1 Basis for model selection and improvement
Traditional occupational-health-risk-assessment methodologies possess notable drawbacks when applied to smart-manufacturing scenarios. The risk-matrix method depends on subjective expert judgment for weight assignment; weight determination lacks objective grounding and cannot capture dynamic shifts as well as coupling interactions among risk factors. The LEC method is only suited for qualitative evaluation within single workplaces and cannot characterize complex cross-scenario and multi-device-coordination-related risks prevalent in intelligent manufacturing [16]. Centered on static indicators, fuzzy comprehensive evaluation struggles to handle risk-event uncertainty and probabilistic reasoning, making real-time risk updating and prediction infeasible.
As a probabilistic-inference paradigm, Bayesian networks excel at processing uncertain information, characterizing factor causal relationships, and implementing dynamic probability updates, enabling faithful representation of risk coupling and stochasticity within smart-production environments. Derived from raw monitoring datasets, the entropy-weighted method calculates indicator weights objectively according to data dispersion degrees, thoroughly eliminating subjective bias originating from manual scoring. Combined with Bayesian networks, it realizes deep integration between objective weighting and dynamic inference [32].
This study fuses the entropy-weighted approach with Bayesian networks to build a dynamic quantitative assessment model. The two components assume well-defined, differentiated roles. The entropy-weighted algorithm computes weight coefficients for each bottom-tier risk indicator from monitoring data and leverages these weights to adjust prior-probability distributions of root nodes inside the Bayesian network. The Bayesian network undertakes risk-factor causal modeling, conditional-probability propagation, and posterior-probability dynamic updating. The conditional-probability table (CPT) is obtained via sample learning or maximum-likelihood estimation; entropy weighting does not participate in conditional-probability computation or posterior-probability updates. This framework achieves intrinsic coupling of objective weighting and dynamic probabilistic reasoning instead of simple serial concatenation of two independent models [33]. It overcomes critical limitations of conventional models including subjective parameter calibration and static analysis and constitutes the core innovation of this paper.
4.2 Construction of an entropy-weighted Bayesian network hybrid model
4.2.1 Calculation of risk indicator weights.
The entropy-weighted method determines weights based on the degree of variation in indicator data; the higher the degree of variation, the stronger the indicator’s discriminatory power in the assessment, and the higher the weight assigned. The steps and mathematical equations for calculating entropy-weighted risk indicators are provided in the “Calculation of Entropy-Weighted Risk Indicators” section of the supporting materials.
Based on the complete formula described above which incorporates standardization, entropy, the coefficient of variation, and normalization, this paper selected 15 core quantitative indicators to calculate their weights, ultimately yielding the entropy-weighted results for each indicator, as shown in Table 2.
The weighted data shows that the error rate in operational instructions, the frequency of information overload, and the duration of gaps in risk monitoring are the top three indicators by weight, each with a weight exceeding 0.100. This indicates that human-machine interaction risks, psychological risks, and management risks are the core dimensions affecting occupational health levels in intelligent manufacturing. Among physiological risks, the concentration of chemical toxin exposure ranked second in weight, indicating that traditional hazard factors still exert a significant influence [33]. Indicators such as the degree of monotony and burnout in work, the frequency of human-machine interface incompatibility, and the coverage rate of occupational health training had lower weights, suggesting that their impact on overall risk is relatively limited.
4.2.2 Building Bayesian network topologies.
Based on the four-level risk indicator system, a directed acyclic graph (DAG) structure is constructed for the Bayesian network. The Bayesian network relies on the directed acyclic graph G = (V, E) and CPT to model risk causal relationships and perform probabilistic inference (S3 Table). For a more detailed discussion of the coupling mechanism between entropy weights and the Bayesian network, see Section 3.2 of the supporting information [34]. In the target layer, intelligent occupational health risks serve as the top-level leaf nodes of the network; in the criterion layer, physiological risks, psychological risks, human-machine interaction risks, secondary environmental risks, and management risks serve as intermediate nodes; and in the factor layer, the 16 core indicators serve as the bottom-level root nodes. The nodes are connected via directed edges, which clarify the causal relationships between risk factors and risk levels. This fully characterizes the risk evolution pathways resulting from the effects of individual factors and the coupling of multiple factors, thereby accommodating the dynamic transmission characteristics of risks in intelligent chemical industries.
This model does not adopt the strong conditional independence assumption of Naive Bayes. The Naive Bayes algorithm requires that all parent nodes be conditionally independent of one another; however, the multilayer Bayesian network constructed in this study follows a general directed acyclic graph (DAG) structure, which allows a single intermediate node to have multiple parent nodes with statistical dependencies [35]. This structure enables the direct encoding of joint interactions among factors through the CPT with multiple parent nodes. Since the risk coupling effect is endogenously characterized through the joint probabilities in the CPT table, there is no inherent conflict between the risk coupling effect emphasized in this paper and the network modeling framework; only the root nodes at the lowest level are assumed to be weakly prior-independent, and this assumption can be overturned through subsequent evidence input and joint inference.
4.2.3 Realization of Bayesian‑network topology and coupling‑effect modeling.
For the directed acyclic graph of the Bayesian network built in this study, the node set is defined as : bottom root‑node set represents the 15 risk indicators; intermediate‑node set corresponds to five risk dimensions, i.e., physiological risk, psychological risk, human‑machine‑interaction risk, secondary‑environmental risk, and management risk; and the top target node denotes the comprehensive occupational‑health‑risk level.
Coupling and nonlinear‑interaction effects across different risk dimensions are endogenously realized by multi‑parent‑node conditional probability tables (CPTs) of intermediate nodes. There is no need to manually add explicit interaction edges or interaction terms in the topology, nor is it necessary to introduce noisy‑OR gates or dynamic‑Bayesian‑network extensions [21]. An intermediate node can simultaneously receive parent‑node inputs from multiple bottom root nodes and other intermediate risk dimensions. The CPT stores conditional probabilities under all combinations of parent‑node states, and naturally characterizes joint dependence and synergistic‑amplification effects when multiple risks occur simultaneously.
4.2.4 Probability parameter calibration.
The model’s probabilistic parameters include prior probabilities and conditional probabilities. Prior probabilities are determined based on on-site monitoring data, historical accident statistics, and industry standard data, with initial values assigned using frequency statistics. Conditional probabilities are derived through maximum likelihood estimation of sample data, combined with the actual operating conditions of intelligent chemical production scenarios, to achieve iterative parameter optimization [36]. All probability parameters have passed consistency tests to ensure that their values conform to the objective laws governing risk evolution. The node information and parameter calculation results for the entropy-weighted Bayesian network model are shown in S2 Table.
4.2.5 Risk‑grade classification criteria and model‑uncertainty analysis.
The output comprehensive occupational‑health risk strictly follows a five‑grade classification scheme: low risk, relatively low risk, moderate risk, relatively high risk, and high risk. The five‑grade classification defined in this subsection applies to model‑output comprehensive occupational‑health risk; an independent sensitivity‑grade classification is additionally defined for characterizing the perturbation capacity of risk factors [37]. The corresponding quantitative risk intervals are 0–0.2, 0.2–0.4, 0.4–0.6, 0.6–0.8, and 0.8–1.0. The classification criteria balance general‑industry applicability and scenario‑specific features and provide unified judgment benchmarks for risk‑assessment outputs.
Three sources contribute to the uncertainties of the Bayesian‑network model for occupational‑health risks in intelligent‑chemical scenarios: ① structural uncertainty originating from indicator‑set selection; ② data uncertainty introduced by noise in on‑site monitoring data; ③ parameter uncertainty resulting from expert judgements during the calibration of prior probabilities and conditional probabilities. Monte‑Carlo probability propagation was adopted to quantify uncertainty propagation. The total output risk variance was decomposed into variances from the three above‑mentioned sources [24,28]:
(2)where = total output risk variance of the model; = structural‑uncertainty variance induced by indicator‑set selection; = data‑uncertainty variance originating from noise in on‑site monitoring measurements; = parameter‑uncertainty variance caused by expert judgement and sample‑learning processes.
For data uncertainty: measurement noise is superimposed on on‑site monitored indicators, and each bottom‑layer indicator obeys a normal distribution:
(3)where is obtained from the manufacturer‑specified measurement error of sensors. Under the Monte‑Carlo framework, 1000 samples were drawn for each bottom‑layer indicator and propagated forward to obtain the risk distribution of the target node [29].
For parameter uncertainty: Dirichlet conjugate priors were applied for sampling‑based perturbation of conditional‑probability‑table parameters, so as to quantify output‑risk fluctuations introduced by expert assignment.
For structural‑indicator‑selection uncertainty: 10% of bottom‑layer indicators were randomly removed for repeated modeling, and the coefficient of variation was used to characterize structural uncertainty:
(4)where is the mean risk score output from repeated modeling runs.
Decomposition results of uncertainty sources are summarized in S3 Table. Calculation results show that parameter uncertainty of conditional‑probability tables derived from expert judgements constitutes the dominant source of model uncertainty, contributing 52.1% of total variance, followed by monitoring‑data noise, whereas structural uncertainty from indicator‑set selection contributes the least.
4.3 Model validation and sensitivity analysis
4.3.1 Model validation.
This study conducted model validation from multiple perspectives, including random partitioning, time series segmentation, multi-model benchmark comparisons, and multi-site simulation cross-validation. The comprehensive results indicate that the entropy-weighted Bayesian network hybrid model demonstrates good predictive capability on the dataset from this site, while also identifying risks of performance degradation caused by time-series drift and distribution shifts across sites.
Time series segmentation validation.
To evaluate the model’s ability to predict the dynamic evolution of risk over time and avoid validation bias caused by random data partitioning, we strictly segmented the data based on the time of sample collection: the first 9 months (241 sample groups) served as the training set, and the last 3 months (85 sample groups) served as the time-series test set. No random shuffling was performed, simulating the real-world engineering scenario of “using historical data to build models and predict future risks over a specific time period.” Evaluation metrics include prediction accuracy, recall, and F1 score.
(5)The validation results for temporal segmentation are shown in S5 Table and S6 Table. After temporal segmentation, the model’s accuracy decreased from 92.3% to 88.2%, indicating a reasonable decline in performance. This was due to a shift in the risk distribution caused by fluctuations in production conditions during the final three months; however, the overall prediction metrics remained at a high level, demonstrating that this hybrid model possesses a certain degree of temporal generalization capability regarding the dynamic evolution of risk and did not exhibit severe temporal overfitting [30]. In practical engineering applications, it is recommended to perform a local update of the model parameters every 6 months using newly acquired monitoring samples to further offset the accuracy loss caused by shifts in the risk distribution.
Multi-model benchmark comparison.
To validate the comprehensive performance of the entropy-weighted Bayesian network hybrid model, we used exactly the same 326 datasets, uniformly split into training and testing sets in a 7:3 ratio, and selected four commonly used risk assessment models of the same type for comparative testing: Independent Bayesian Network (without entropy weighting; original equal-weighted root nodes), Fuzzy Comprehensive Evaluation, Random Forest (RF), and XGBoost gradient boosting trees. Evaluation metrics included prediction accuracy, recall, score, and AUC; the comparison results are shown in Table 3.
The results of the benchmark comparison show that:
- 1) The performance of the independent Bayesian network declines significantly when entropy-weighted objective weights are not introduced, demonstrating that the fusion strategy of modifying the root node’s prior probability using entropy weights can effectively improve the evaluation performance of the Bayesian network;
- 2) Machine learning algorithms such as RF and XGBoost exhibit high predictive accuracy but are black-box models; they struggle to provide causal explanations for relationships among risk factors and are unable to analyze risk causality propagation or coupling mechanisms;
- 3) The fuzzy comprehensive evaluation method demonstrated the lowest overall performance and was significantly influenced by the settings of the subjective membership function;
- 4) The entropy-weighted Bayesian hybrid model proposed in this paper achieved the best performance across all quantitative metrics. It retains the causal interpretability advantage of probabilistic graphical models while balancing predictive accuracy with the ability to explain risk mechanisms, making it better suited to the research needs of occupational health risk assessment.
Multi-site simulation cross-validation.
Since it was not possible to obtain complete, matched sets of real-world measured data from multiple chemical plants, this study used Monte Carlo simulation to construct three virtual chemical plant datasets with distinct distributions, corresponding to high-, medium-, and low-intelligence chemical plants (operating conditions A/B/C, consistent with the simulation conditions described in Section 4.3.2) [25]. Five hundred sets of simulated samples were generated for each site. The training set used only samples from the real-world site (Zhenhai Refining & Chemical), while testing was conducted directly on the three simulated sites to perform cross-site simulation cross-validation, thereby evaluating the model’s out-of-data generalization capability and identifying the risk of overfitting.
The results of the multi-site simulation cross-validation are shown in S5 Table and S6 Table. The results indicate that when the intelligence level of the test sites is close to that of the real-world site in the training set, the model maintains high accuracy; as the level of automation decreases, leading to missing monitoring indicators and shifts in the distribution of risk factors, the model’s performance gradually declines. This phenomenon is not due to overfitting but rather results from shifts in risk distribution across facilities [37]. The simulation results suggest that, to transfer the model to a new plant, it is recommended to fine-tune the prior probabilities of the underlying root nodes using a small number of local samples.
4.3.2 Simulation testing of the model’s cross-facility generalization capability.
Three simulation scenarios were set up to model chemical plants with different levels of automation:
- Scenario A (High Automation, benchmarked against Zhenhai Refining & Chemical): 75% of positions are remotely monitored, with 94% sensor coverage;
- Scenario B (Medium Automation): 45% of positions are remotely monitored, with 65% sensor coverage;
- Scenario C (Low Automation): 20% of positions are remotely monitored, with 40% sensor coverage.
Monte Carlo simulation generated 500 sample sets for each operating condition, which were input into the trained entropy-weighted Bayesian network to produce the results shown in Table 4.
Simulation results show that the model for highly intelligent factories achieves the highest accuracy; as the level of intelligence decreases, the model’s accuracy gradually declines. This performance degradation stems primarily from shifts in the distribution of job roles and missing monitoring indicators. Two engineering improvement strategies are proposed: (1) Fine-tune the conditional probability table using transfer learning, which can improve simulation accuracy under operating condition C to 83.4%; (2) Enable missing evidence reasoning to prevent samples from being discarded solely due to missing indicators.
4.4 Sensitivity analysis
Sensitivity analysis is used to identify key controlling factors that influence risk levels. By altering the probability of the root node using the single-factor perturbation method, the magnitude of change in the posterior probability of the top-level target node is calculated; a larger magnitude of change indicates higher sensitivity of the indicator. In this study, 15 core indicators were selected for sensitivity analysis, and sensitivity coefficients were calculated and ranked. Specific data are shown in Fig 5.
Sensitivity analysis indicates that the error rate in operational instructions, the duration of gaps in risk monitoring, and the frequency of information overload are factors with an extremely high sensitivity level and are the core drivers of rising occupational health risks in the intelligent chemical industry [26]. Factors such as chemical toxin exposure concentrations and anxiety levels regarding operational responsibilities are classified as high-sensitivity factors and have a significant impact on risk levels. Factors such as high temperature and high humidity parameters and training coverage rates have a lower sensitivity level and exert a limited influence on overall risk.
In addition to single-factor sensitivity analysis, this study conducted multiple sets of two-factor coupled perturbation simulations; the simulation results are shown in S4 Table. It can be observed that a weak inhibitory effect (Ωij < 1) occurs between some traditional physiological and environmental risks; not all risks are amplified. The model can reproduce three types of coupling behaviors amplification, linear superposition, and inhibition demonstrating that the CPT conditional probability table has the capability to characterize diverse coupling interactions. Bayesian networks rely on the conditional probability table within multi-parent nodes to endogenously capture risk coupling interactions, eliminating the need to manually add interaction edges or implement additional model extensions; As a post-processing tool, it is used to analyze the coupling strength of outputs from a fully trained network [27]. Simulation results indicate that new risks in the human-computer interaction and psychological categories generally exhibit positive amplification; certain combinations of traditional physiological and environmental risks exhibit a weak suppression effect, demonstrating the model’s ability to reproduce diverse coupling behaviors.
5. Empirical analysis and results
5.1 Case studies and data collection
Ningbo Zhenhai Refining & Chemical is a large-scale intelligent petrochemical production base in China. It has completed digital transformations across four key areas intelligent control, intelligent inspection, intelligent warehousing, and intelligent operations and maintenance and is equipped with an industrial internet platform, intelligent sensing devices, remote control systems, and robotic work units [13]. The facility exhibits the typical characteristics of intelligent chemical production and is well-suited to the empirical research requirements of this study. This study was conducted as part of a collaborative project with the Zhenhai Refining & Chemical Branch of China Petroleum & Chemical Corporation (Sinopec). Researchers entered the plant premises to conduct fieldwork in their capacity as technical staff of the collaborating project, without the need for an independent on-site inspection permit issued by external government authorities. All employee interviews, questionnaire surveys, and on-site data collection were carried out after the collaborating enterprise granted approval for access to the plant premises [15]. Data collection spanned 12 months and covered different production cycles and work shifts. Data sources included on-site environmental monitoring data, employee occupational health examination reports, equipment operating parameters, safety management records, employee interview transcripts, and accident statistics. A total of 326 valid data sets were collected, encompassing quantitative data for 15 core risk indicators, providing ample data support for the empirical analysis.
5.2 Risk identification results and statistical analysis
The results of risk identification in the empirical scenarios show that risk factors in the intelligent control scenario accounted for the highest proportion, reaching 38.2%, primarily due to information overload and command errors during remote operation. The intelligent inspection scenario accounted for 27.5%, with risks primarily related to human-machine collaboration and traditional physical hazards. The intelligent warehousing scenario accounted for 19.3%, with risks concentrated on abnormal environmental parameters and hidden hazards in equipment operation [16]. The intelligent operation and maintenance scenario accounted for 15.0%, with risks primarily stemming from management loopholes and emergency response deficiencies.
High-frequency risk factors included operational command errors, information overload, and lack of risk monitoring, which collectively accounted for over 50%. Latent risks accounted for 37.6%, primarily including psychological stress, physical fatigue, and latent equipment failures risk types that are difficult to capture using traditional identification methods [17]. This confirms the highly concealed nature of risks in intelligent chemical plants and validates the effectiveness of the hybrid identification method used in this study.
5.3 Results of quantitative risk level assessment
The model’s quantitative assessment results show that the overall occupational health risk quantification value for the pilot site is 0.726, indicating a high-risk level. The ranking of risk weights across dimensions is as follows: human-machine interaction risks, management risks, psychological risks, physiological risks, and secondary environmental risks [9]. Risks related to human-machine interaction and management control have emerged as the core issues.
A total of eight high-risk nodes were identified: the remote control console, the intelligent inspection robot work area, the central monitoring room, the reactor intelligent control unit, the intelligent material conveyance node, the ventilation control module, the risk monitoring terminal, and the emergency response control console. The results of the risk coupling probability analysis indicate that the probability of multi-factor coupled risks occurring is 42.3%, significantly higher than that of single risks; the coupling effect is the core factor exacerbating risk hazards.
5.4 Sensitivity analysis results
The results of the empirical sensitivity analysis are consistent with the theoretical model analysis. The error rate of operating instructions, the duration of risk monitoring gaps, and the frequency of information overload are the core controlling factors, with sensitivity coefficients all exceeding 0.240. A 10% disturbance in any of the above factors will cause the overall risk level to fluctuate by more than 20%, making them core targets for risk management. The sensitivity of psychological and management risk factors is generally higher than that of physiological and secondary environmental risks, indicating that after the transition to intelligent systems, the impact of new derivative risks has surpassed that of traditional residual risks.
5.5 Closed‑loop intervention and effect verification
Based on the results of sensitivity analysis, risk‑sensitivity values only quantify the magnitude of factor impacts on risk outputs. Highly sensitive risk factors do not necessarily correspond to the optimal intervention targets in practical engineering scenarios. Intervention priorities should also consider practical constraints including implementation cost, operability, reversibility, and stakeholder acceptance. Accordingly, this study adopted the entropy‑weighted TOPSIS multi‑criteria decision‑making (MCDM) framework to rank intervention‑target priorities for targeted closed‑loop intervention. Interrupted time‑series (ITS) analysis was performed on monthly on‑site monitoring time‑series datasets to characterize risk evolution before and after intervention [5–9].
Priority evaluation of intervention targets using multi‑criteria decision‑making.
A multi‑criteria evaluation system was established with five assessment criteria:
: Risk sensitivity (derived from sensitivity coefficients; higher values indicate greater risk impacts);: Intervention implementation cost (1‑5 scoring scale; lower scores represent higher transformation costs);: Reversibility of intervention measures (1‑5 scoring scale; higher scores indicate easier adjustment or revocation of measures);: On‑site operation and maintenance practicality (1‑5 scoring scale; higher scores represent lower enterprise implementation difficulty);: Stakeholder acceptance (1‑5 scoring scale; higher scores mean less resistance from operation and maintenance staff and better acceptance).
Detailed methodologies and mathematical formulations for multi‑criteria priority evaluation are provided in the Supporting Information (Part 4).
Table 5 summarizes multi-criteria TOPSIS prioritization results for intervention targets. Key observations are as follows:
- The three highest-sensitivity factors operational-command error rate, risk-monitoring-gap duration, and information-overload frequency also yield the highest overall proximity values. Considering both risk impacts and real-world engineering constraints, these factors are designated as first-tier intervention priorities.
- Chemical-exposure concentration ranks fourth in risk sensitivity. Nevertheless, high implementation costs, poor reversibility, and substantial on-site deployment challenges reduce its comprehensive priority to sixth. This finding demonstrates that high-sensitivity factors do not automatically constitute optimal intervention targets under practical engineering conditions; multi-criteria constraints including cost and feasibility must be incorporated to adjust intervention sequencing.
A two-tier intervention implementation strategy is established based on MCDM-derived comprehensive priority rankings:
- First Tier (Priority Implementation): operational-command error rate, risk-monitoring-gap duration, information-overload frequency. Corresponding measures include human-machine-interface optimization, alarm-signal de-noising, and monitoring-system upgrading.
- Second Tier (Medium-to-Long-Term Gradual Implementation): operational-responsibility-anxiety level, equipment-response-latency duration, chemical-exposure concentration. Corresponding measures include employee psychological-counseling services, iterative equipment retrofitting, and on-site process-containment modifications.
Fig 6 and S7 Table present pre- versus post-intervention risk-level comparisons. Following intervention, risk values for all monitoring nodes and core indicators exhibit statistically significant downward trends. Overall risk levels decline from high-risk and moderately-high-risk categories to moderate-risk and low-risk categories. Core sensitive factors show the most pronounced reduction: risk-reduction magnitudes for operational-command error rate, risk-monitoring-gap duration, and information-overload frequency all exceed 60%, far outperforming other indicators. Risk reduction for high-risk nodes across individual scenarios ranges from 40% to 50%, bringing aggregate risk under effective control.
The substantial risk reduction stems from interventions that precisely target high-sensitivity key control factors and block risk-transmission pathways at their sources. Human-machine-interface optimization directly lowers operational-command error rates and equipment-response latency, mitigating human-machine-interaction hazards. Upgraded real-time risk-monitoring systems eliminate monitoring gaps and enable early risk alerts together with dynamic risk governance. Psychological counseling and on-the-job training effectively alleviate information overload and operational anxiety and decrease psychological-risk magnitudes. Conventional hazard-mitigation measures strengthen management over chemical-toxicant exposure, noise, and vibration and restrain further deterioration of physiological risks.
This study presents a tiered closed-loop-intervention-system framework diagram (Fig 7), which illustrates the full-chain intervention logic covering source control, process governance, endpoint protection, and optimization feedback and demonstrates the engineering application value of this research. Synergistic implementation of multiple measures mitigates risk-coupling effects, lowers multi-factor compound-risk probabilities, and drives sustained overall-risk decline. Empirical outcomes fully verify that entropy-weighted-Bayesian-network-driven dynamic assessment plus targeted interventions deliver robust risk-governance performance and substantially improve occupational-health protection within intelligent chemical-production environments.
6. Discussion
6.1 Key study findings
Against the backdrop of intelligent-transformation-driven occupational-health-governance challenges within the chemical sector, this study deconstructs four typical intelligent chemical-production operational scenarios, develops a four-tier risk-indicator system, and builds a hybrid dynamic entropy-weighted Bayesian-network assessment model. Empirical testing and closed-loop intervention validation are completed using the Zhenhai Refining and Chemical production base as the research site. Core findings are summarized below.
First, occupational-health-risk landscapes undergo structural shifts under intelligent chemical-production conditions. While conventional physicochemical hazards such as chemical-toxicant exposure and noise are suppressed, human-machine-interaction, psychological, and managerial risks emerge as dominant hazard categories. Emerging derivative risks including information overload, operational-command errors, and physiological strain induced by prolonged screen-based remote-console work represent non-negligible hazards for intelligent chemical-manufacturing scenarios. These outcomes align with qualitative conclusions reported in existing literature; the present study further supplies quantitative empirical evidence supported by 15 measurable indicators and field-monitoring datasets.
Second, risk-coupling effects are prevalent within intelligent chemical-production contexts and can substantially amplify occupational-health hazards. This study employs multi-parent-node conditional-probability tables (CPT) from Bayesian-network modeling to endogenously characterize cross-dimensional risk-coupling behaviors. Post-processing simulation for the coupling coefficient reveals that couplings between human-machine-interaction and psychological risks mostly produce strong positive amplification (). In contrast, select conventional physiological-environmental risk combinations yield weak inhibitory effects (). This observation indicates that risk coupling does not universally exacerbate harm; diverse interaction modes exist, and co-occurrence of multiple risks cannot be assumed to automatically amplify adverse consequences.
Third, the hybrid entropy-weighted Bayesian-network model exhibits favorable comprehensive performance for occupational-health-risk assessment in intelligent-chemical-production settings. Under time-series-segmentation validation, the model achieves 88.2% prediction accuracy and outperforms competing approaches including standalone Bayesian networks and fuzzy comprehensive evaluation. Compared with machine-learning algorithms such as RF and XGBoost, the proposed model balances prediction accuracy and causal interpretability. It not only outputs quantitative risk levels but also analyzes coupling and transmission relationships among risk factors, making it better suited for safety-assessment applications requiring mechanistic explanations such as occupational-health-risk evaluation. Nevertheless, simulation results demonstrate that prediction performance degrades for facilities with low intelligence levels due to incomplete monitoring indicators and risk-distribution drift, which suggests that parameter fine-tuning is required for cross-enterprise model migration.
Fourth, risk-sensitivity coefficients alone are insufficient for determining practical intervention priorities in engineering practice. A high-sensitivity factor only reflects its mathematical influence on risk outputs; practical engineering constraints including intervention cost, reversibility, operational-maintenance feasibility, and stakeholder acceptance must also be incorporated. Entropy-weighted TOPSIS multi-criteria decision-making is adopted to rank intervention targets. Results indicate that despite high sensitivity, chemical-exposure concentration receives a greatly reduced comprehensive intervention priority owing to high retrofitting costs. Interrupted time-series (ITS) analysis (Supporting Information Part 5, S6 Table) demonstrates statistically significant step-wise and sustained overall-risk decline after tiered closed-loop intervention implementation. Limited by field conditions, no blank control group is available; thus observed risk reduction represents combined effects of multiple contributing factors and cannot be attributed exclusively to the deployed intervention program.
6.2 Comparisons with domestic and international published research
Risk-identification and indicator-system domain.
Hassim et al. performed pioneering occupational-health-risk-identification research targeting chemical-industry facilities [13]. Nevertheless, their work mainly focuses on conventional chemical-process plants; indicator systems emphasize on-site physicochemical exposures such as toxic-substance and noise hazards while giving limited attention to human-machine-related and psychological risks originating from intelligent upgrades. Several domestic studies investigating smart-chemical-manufacturing occupational health have identified emerging concerns such as remote-operation-induced fatigue and cognitive overload. Most existing work remains confined to qualitative description with fragmented indicator dimensions, fails to differentiate divergent risk features across distinct smart-operation scenarios, and lacks comprehensive hierarchical quantitative indicator frameworks [37].
Relative to prior literature, this study presents three distinct contributions:
- Four intelligent-chemical-production scenarios are classified according to business functions rather than simple physical-plant geography, clarifying fundamental disparities in risk triggers across scenarios.
- A four-tier progressive indicator system is constructed that integrates conventional residual hazards, emerging derivative risks, as well as coupled and cumulative risks. Latent risks such as psychological stress, information overload, and human-machine-interface defects are rendered quantifiable.
- 3. Indicator-perturbation experiments are conducted to quantitatively evaluate outcome variations caused by indicator substitution, supplying actionable reference for indicator replacement for enterprises constrained by monitoring resources [38].
Risk-assessment-model methodological domain.
Conventional chemical-industry occupational-health-risk assessment widely adopts risk-matrix, LEC, and fuzzy-comprehensive-evaluation approaches. These static methodologies depend heavily on subjective expert scoring and struggle to characterize risk dynamic evolution and inter-factor coupling relationships [7–9]. Multiple prior publications implement standard Bayesian networks for safety-risk evaluation; however, most assign all prior probabilities based purely on expert judgment and introduce substantial subjective bias. Machine-learning-driven approaches (RF, XGBoost) can achieve high prediction accuracy yet operate as black-box models that hinder mechanistic interpretation of risk causal and coupling pathways [12].
Lu et al. applied entropy-weighted Bayesian modeling for ecological-environmental-risk evaluation [31]. Their workflow modifies conditional probabilities via entropy weighting, which risks distorting causal relationships learned from empirical datasets. The model developed herein introduces a key distinction: entropy weighting adjusts only prior probabilities of bottom-layer root nodes while leaving conditional-probability tables unmodified, thus preserving data-derived risk-causal logic [31,32]. This design leverages monitoring datasets for objective weighting and avoids distortion of inherent factor-coupling relationships. Furthermore, this study avoids direct adoption of the noise-OR gate assumption. Multi-parent-node CPTs endogenously model coupling behaviors, and post-processing calculates coupling-amplification coefficients to quantify pairwise-factor-coupling strength. This realizes a combined paradigm of “in-model endogenous-coupling simulation plus external coupling-strength quantification” and addresses limitations of earlier research that only provided qualitative coupling descriptions [33].
Comparisons with HCPS, control-room ergonomics, and process-AI literature.
Rooted in Healthy Operator 4.0 and human-cyber-physical-system (HCPS) theoretical foundations, this research complements three streams of cutting-edge published work.
First, existing HCPS / Operator 5.0-oriented studies predominantly target discrete manufacturing. This investigation deconstructs four chemical-process-industry intelligent-operation scenarios and translates HCPS human-machine-coupling characteristics into 15 quantifiable occupational-health indicators, filling gaps of practical indicator frameworks for process-industry applications [10].
Second, existing control-room cognitive-ergonomics research mostly relies on simulation experiments and focuses on short-duration human errors and process-safety incidents. The present study treats cognitive-level information overload and command errors as mediating variables and further links them to long-term chronic physiological and psychological occupational-health harms. It establishes the transmission chain: human-machine-interaction characteristics → cognitive load → occupational-health damage, expanding cognitive-ergonomics research scope from process safety toward occupational-health domains [23].
Third, most process-industry AI-driven risk-prediction work uses accident occurrence as prediction targets, and black-box models obstruct mechanistic interpretation. This paper adopts a hybrid entropy-weighted Bayesian-network model that leverages real-world field-monitoring data for quantitative risk assessment while retaining causal interpretability inherent to probabilistic-graphical models [27]. It outputs risk magnitudes and quantitatively analyzes risk-coupling effects including amplification or inhibition between factors and delivers mechanistic support for targeted interventions, mitigating interpretability deficiencies of purely data-driven AI approaches [15].
Nevertheless, this study has not incorporated multimodal physiological-sensor datasets such as eye-tracking and electroencephalogram (EEG) signals to realize in-situ real-time cognitive-load measurement. Follow-up work may integrate wearable-sensor technologies and cognitive-ergonomics experimental protocols to further enrich model input sources.
6.3 Study strengths and limitations
Strengths.
- Scenario-decomposition perspective: Four intelligent-manufacturing scenarios are defined based on business functions. Conceptual boundaries between scenarios are clarified, supporting differentiated scenario-targeted governance.
- Methodological innovations: The hybrid entropy-weighted Bayesian-network model integrates data-driven objective weighting and probabilistic causal inference. This work explicitly specifies that entropy weighting only revises root-node prior probabilities while sample-learned conditional-causal relationships remain intact. It also distinguishes endogenous CPT-based coupling simulation from post-processing coupling-coefficient analysis.
- Empirical rigor: Multiple technical measures are deployed to mitigate subjective bias across evaluation and empirical analysis, including time-series-segmentation validation, multi-model benchmark comparison, Monte-Carlo uncertainty decomposition, entropy-weighted-TOPSIS intervention-target ranking, and interrupted time-series evaluation.
- Engineering practicality: Research outputs including indicator systems, model parameters, and intervention-priority lists can be directly reused by petrochemical enterprises and regulatory authorities and possess tangible implementation value.
Limitations
- Single empirical-data source: All empirical datasets originate from one large, highly-intelligent petrochemical complex (Zhenhai Refining and Chemical Plant). Although multi-condition simulations are performed for medium- and low-intelligence factories, measured field data from multiple enterprises spanning varying intelligence maturity levels are absent. Further field-dataset validation is required to verify external generalization performance.
- Coupling-analysis constraints: Current post-processed coupling-amplification coefficients only implement perturbation analysis for pairwise-factor interactions. High-order multidimensional couplings involving three or more factors remain incompletely explored; quantitative characterization of high-order couplings constitutes a remaining model limitation.
- Data-acquisition constraints: Select psychological indicators (operational-responsibility anxiety, job burnout) depend on subjective questionnaire scales and cannot be fully automatically captured via sensors, introducing potential self-report bias.
- Temporal-model constraints: The current model represents a static Bayesian network. Temporal risk evolution is realized via post-processing ITS analysis; native dynamic Bayesian-network (DBN) architecture that continuously characterizes time-variant risk trajectories has not been constructed.
- Self-report-related reporting bias: Partial psychological-dimension indicators (task-related anxiety, monotony-driven burnout) derive from employee questionnaire self-reports. Such data inevitably contain social-desirability bias and recall bias; subjective scores may also be influenced by respondent mood and on-site workplace atmosphere.
Conclusion
Targeting practical challenges including occupational-health-risk identification obstacles, assessment-model subjectivity, and insufficient multi-dimensional quantitative support for intervention targets under intelligent chemical-production transformation, this study deconstructs four business-function-oriented intelligent chemical-production scenarios. A four-tier progressive occupational-health-risk indicator system covering objective, criterion, factor, and indicator layers is constructed. Human-machine-interaction risks, psychological risks, and managerial risks are identified as dominant hazard categories for modern intelligent-chemical-production environments.
A hybrid entropy-weighted Bayesian-network assessment model is developed. Model validation is performed via time-series segmentation, multi-model benchmarking, and Monte-Carlo uncertainty decomposition. The model attains 88.2% prediction accuracy on time-series test datasets, and three high-sensitivity key risk drivers are pinpointed: operational-command error rate, information-overload frequency, and risk-monitoring-gap duration. Empirical analysis is performed using the Zhenhai Refining and Chemical base, where overall occupational-health risks are found to reside at a relatively high level. An entropy-weighted TOPSIS multi-criteria decision-making framework is introduced. Intervention priorities are determined by comprehensively weighing risk-sensitivity, implementation cost, and practical feasibility, overcoming drawbacks of priority-setting that relies solely on sensitivity metrics. A six-month tiered closed-loop intervention is implemented, and interrupted time-series analysis confirms statistically significant step-wise risk reduction.
This research delivers theoretical tools and practical operational guidance for petrochemical-enterprise occupational-health-risk governance within smart-manufacturing contexts. Nevertheless, this investigation possesses acknowledged limitations: single-site empirical data, self-report bias within psychological questionnaires, static entropy-weight computation, and limited bidirectional-feedback representation within directed-acyclic-graph topology. Future research will pursue multi-site field validation, integrate multi-source objective physiological datasets, adopt sliding-window-based dynamic entropy weighting, and extend toward dynamic Bayesian-network architectures to further improve model generalization capability and real-time assessment performance.
Supporting information
S2 Fig. Causal transmission logic model illustrating differentiated occupational health risks across four types of smart chemical production scenarios.
https://doi.org/10.1371/journal.pone.0363258.s002
(TIF)
S1 Table. Comparison of risk characteristics between traditional and intelligent chemical production scenarios.
https://doi.org/10.1371/journal.pone.0363258.s003
(DOCX)
S2 Table. Summary of test data for indicator‑replacement perturbation experiments.
https://doi.org/10.1371/journal.pone.0363258.s004
(DOCX)
S3 Table. Decomposition results of uncertainty sources.
https://doi.org/10.1371/journal.pone.0363258.s005
(DOCX)
S4 Table. Simulation results of pairwise coupling‑amplification coefficients for typical risk‑factor pairs.
https://doi.org/10.1371/journal.pone.0363258.s006
(DOCX)
S5 Table. Validation results under time‑series data splitting.
https://doi.org/10.1371/journal.pone.0363258.s007
(DOCX)
S6 Table. Regression estimation outputs of the interrupted‑time‑series (ITS) model.
https://doi.org/10.1371/journal.pone.0363258.s008
(DOCX)
S7 Table. Comparison of risk levels before and after intervention.
https://doi.org/10.1371/journal.pone.0363258.s009
(DOCX)
S1 File. Model construction and computational process.
https://doi.org/10.1371/journal.pone.0363258.s010
(DOCX)
Acknowledgments
The authors would like to express their sincere gratitude to all individuals and institutions who contributed to the completion of this study. Special thanks are extended to the data providers, peer reviewers, and colleagues who offered valuable insights and suggestions throughout the research process. Generative-AI Usage Statement: During the preparation of this manuscript, generative artificial intelligence tools were used solely for English language polishing, sentence structure optimization, standardization of charts and tables, and reference to literature. All core research concepts, scenario classification, indicator system construction, model design, on-site data collection, experimental calculations, result analysis, and the writing of the main body of the manuscript were independently completed by the authors. Artificial intelligence was not involved in the formulation of research hypotheses, the generation of raw data, the calibration of model parameters, the derivation of research conclusions, or the creation of scientific content. The authors have thoroughly reviewed and revised all text processed by AI and assume full responsibility for the completeness, accuracy, and validity of the entire manuscript.
References
- 1. Yang S, Navarathna P, Ghosh S, Bequette BW. Hybrid modeling in the era of smart manufacturing. Comp Chem Eng. 2020;140:106874.
- 2. Edgar TF, Pistikopoulos EN. Smart manufacturing and energy systems. Comp Chem Eng. 2018;114:130–44.
- 3. Sagdic K, Eş I, Sitti M, Inci F. Smart materials: rational design in biosystems via artificial intelligence. Trends Biotechnol. 2022;40(8):987–1003. pmid:35241311
- 4.
Kusiak A. Smart manufacturing. In: Springer handbook of automation. 973–85). Cham: Springer International Publishing; 2023.
- 5. Aldrini J, Chihi I, Sidhom L. Fault diagnosis and self-healing for smart manufacturing: A review. J Intell Manuf. 2023;35(6):2441–73.
- 6. Hovanec M, Korba P, Vencel M, Al-Rabeei S. Simulating a digital factory and improving production efficiency by using virtual reality technology. Appl Sci. 2023;13(8):5118.
- 7. Huang L, Cai G, Yuan H, Chen J. A hybrid approach for identifying the structure of a Bayesian network model. Expert Syst Appl. 2019;131:308–20.
- 8. Cieply J, Disselkamp J-P, Dyck F, Alturk W, Kühn Dr-IngA, Dumitrescu ProfDr-IngR. A systematic literature review on the Digital Twin of the factory in the context of the product and factory lifecycle. Procedia CIRP. 2023;119:834–40.
- 9.
Altenried M. The digital factory: The human labor of automation. The Digital Factory. University of Chicago Press; 2022.
- 10. Bellavista P, Bicocchi N, Fogli M, Giannelli C, Mamei M, Picone M. Requirements and design patterns for adaptive, autonomous, and context-aware digital twins in industry 4.0 digital factories. Comp Industry. 2023;149:103918.
- 11. Costa O, Matias J, Pimentel C. Occupational health and safety in construction projects: A case study on chemical industry sector. IJOOES. 2021;5(2):14–21.
- 12. Swaminathan V. Occupational health and safety in chemical industries in transitional economies. Indian J Occup Environ Med. 2011;15(3):85–6.
- 13. Hassim MH. Comparison of methods for assessing occupational health hazards in chemical process development and design phases. Curr Opin Chem Eng. 2016;14:137–49.
- 14. Schulte PA, McKernan LT, Heidel DS, Okun AH, Dotson GS, Lentz TJ, et al. Occupational safety and health, green chemistry, and sustainability: A review of areas of convergence. Environ Health. 2013;12:31. pmid:23587312
- 15. Ott MG, Oberlinner C, Lang S, Hoffmann G, Nasterlack M, Pluto R-P, et al. Health and safety protection for chemical industry employees in a rotating shift system: Program design and acute injury and illness experience at work. J Occup Environ Med. 2009;51(2):221–31. pmid:19209044
- 16. Karaboyaci M, Kandemir H, Uysal E. Measurement of work safety and occupational health perceptions of chemical sector employees. Fresenius Environ Bulletin. 2019;28(11):8511–9.
- 17. Hassim MH, Edwards DW. Development of a methodology for assessing inherent occupational health hazards. Process Saf Environ Prot. 2006;84(5):378–90.
- 18. Derakhshan Jazari M, Soltanzadeh A, Heidari H, Mohammadi H, Mohammadbeygi A. Modeling causal factors of occupational accidents in chemical industries: A 10-year field study in Iran. Iranian J Chem Chem Eng. 2021;40(1):357–65.
- 19. Akash MAS, Sah BP. Smart occupational safety management through IOT sensor networks, machine learning, and real-time risk assessment in chemical processing plants. ASRC Procedis: Glob Presp Sci and Scholar. 2025;01(01):1958–98.
- 20. Rout BK, Sikdar BK. Hazard identification, risk assessment, and control measures as an effective tool of occupational health assessment of hazardous process in an iron ore pelletizing industry. Indian J Occup Environ Med. 2017;21(2):56–76.
- 21. Adikwu FE, Ozobu CO, Odujobi O, Onyekwe FO, Nwulu EO. A comprehensive review of Health Risk Assessments (HRAs) and their impact on occupational health programs in large-scale manufacturing plants. Int J Multidisp Growth Eval. 2025;6(1):1525–38.
- 22. Raja U, Iqbal N. Ensuring worker safety in Lahore’s large industries: a study on occupational health, safety, and risk management. J Energy Environ Policy Options. 2019;2(4):117–24.
- 23. Damilos S, Saliakas S, Karasavvas D, Koumoulos EP. An overview of tools and challenges for safety evaluation and exposure assessment in industry 4.0. Appl Sci. 2024;14(10):4207.
- 24. Pishgar M, Issa SF, Sietsema M, Pratap P, Darabi H. REDECA: A novel framework to review artificial intelligence and its applications in occupational safety and health. Int J Environ Res Public Health. 2021;18(13):6705. pmid:34206378
- 25.
Ozobu CO, Adikwu FE, Odujobi O, Onyekwe FO, Nwulu EO, Daraojimba AI. Leveraging AI and machine learning to predict occupational diseases: A conceptual framework for proactive health risk management in high-risk industries. 2023.
- 26. Lemos J, de Souza VB, Falcetta FS, de Almeida FK, Lima TM, Gaspar PD. A System for individual environmental risk assessment and management with IoT based on the worker’s health history. Appl Sci. 2024;14(3):1021.
- 27.
Sánchez Segura MI, Dugarte Peña GL, Amescua Seco AD, Medina Domínguez F. Smart occupational health and safety for a digital era and its place in smart and sustainable cities. 2021.
- 28. Leso V, Fontana L, Iavicoli I. The occupational health and safety dimension of Industry 4.0. La Medicina del lavoro. 2018;109(5):327.
- 29. Ozobu CO. Modeling exposure risk dynamics in fertilizer production plants using multi-parameter surveillance frameworks. Iconic Res Eng J. 2020; 4(2), 227–32.
- 30. Simeone A, Grant R, Ye W, Caggiano A. A human-cyber-physical system for Operator 5.0 smart risk assessment. Int J Adv Manuf Technol. 2023;129(5–6):2763–82.
- 31. Lu X, Guo S, Zhang D, Sun J, Shi W. Development of an eco-environmental evaluation system for islands in jiangsu, china, based on the time-varying entropy weight method and a bayesian network. Sustainability. 2026;18(10):4769.
- 32.
Ma J, Zhou Z, Jiang Y, Ao J, Wan H. A sparse bayesian entropy-weighted cloud model for fault diagnosis and self-healing in distribution networks. In: 2025 4th International Conference on Renewable Energy and Electrical Technology (ICREET). 2025. 101–5. https://doi.org/10.1109/icreet68688.2025.00026
- 33. Rathnakumar R, Lu X, Huang J, Yan H, Liu Y. Bayesian entropy neural networks for physics-aware prediction. Reliability Eng System Safety. 2026;266:111575.
- 34. Zheng H, Tang Y. Deng entropy weighted risk priority number model for failure mode and effects analysis. Entropy (Basel). 2020;22(3):280. pmid:33286052
- 35. Fan Z, Xu F. Health risks of occupational exposure to toxic chemicals in coal mine workplaces based on risk assessment mathematical model based on deep learning. Environ Tech Innov. 2021;22:101500.
- 36. Achumie GO, Oyegbade IK, Igwe AN, Ofodile OC, Azubuike C. A conceptual model for reducing occupational exposure risks in high-risk manufacturing and petrochemical industries through industrial hygiene practices. Indus Hygiene Safety J. 2022;1(1):26–37.
- 37. Wang J, Lee J-S, Amouri REL, Ip JC-H, Liu W, Mo J. Reproductive toxicity of benzo[α]pyrene: Effects, mechanisms, and perspectives. Comp Biochem Physiol C Toxicol Pharmacol. 2026;308:110608. pmid:42398748
- 38. Esmaeili SV, Alboghobeish A, Yazdani H, Ghasemi Koozekonan A, Pouyakian M. Optimizing in-store warehouse safety: A DEMATEL approach to comprehensive risk assessment. PLoS One. 2025;20(2):e0317787. pmid:39946382