TL;DR
This paper uses Bayesian Networks (BNs) to quantitatively model the complex relationship between worker states (like fatigue, illness, or lack of experience) and unsafe behaviors in coal mine accidents. By developing a novel fuzzy probability assessment method and employing expert knowledge, they identify that poor worker states significantly increase the likelihood of violations and decision-making errors, with insufficient experience and poor fitness for duty being key contributing factors.
What Problem Does This Paper Address?
Coal mine accidents remain a significant concern globally, with human error consistently identified as a primary cause. While accident investigations often point to unsafe behaviors like rule violations, there's a recognized need to understand the underlying worker states that contribute to these behaviors. Factors such as mental fatigue, illness, and lack of experience are frequently cited but not always systematically quantified in their impact on accident causation. This paper aims to bridge this gap by providing a quantitative analysis of how various worker states influence specific unsafe behaviors in the context of Chinese coal mines, a high-risk industry.
Methods
The core of this research lies in the application and extension of Bayesian Networks (BNs). The authors first construct a BN structure based on the HFACS (Human Factors Analysis and Classification System) framework and expert knowledge from three coal mine safety bureau experts. This structure represents causal relationships between identified worker states and unsafe behaviors.
A key methodological contribution is the development of a verbal versus numerical fuzzy probability assessment method. This approach is designed to elicit conditional probabilities from experts when large datasets are unavailable, a common challenge in accident investigation data. It combines linguistic variables (e.g., "certain," "probable") with triangular fuzzy numbers to represent expert uncertainty. The junction tree algorithm is then used for inference within the BN.
Datasets
The study primarily relies on expert knowledge elicited through interviews and accident investigation reports. Specifically, they analyzed 163 accident investigation reports from the Fenxi Coal Mine Safety Bureau in Shanxi, China, to identify direct and indirect worker states related to unsafe behavior. Additionally, three experts from the Fenxi Coal Mine Safety Bureau were consulted to define relationships and probabilities within the BN. No new benchmark datasets were introduced; instead, existing accident investigation data and expert opinions were synthesized.
Key Insights and Ideas
- Worker State is a Critical Predictor: The research strongly reinforces the idea that a worker's internal state is not just a background factor but a direct and significant driver of unsafe actions.
- Unsafe Behaviors Aren't Equal: Not all unsafe behaviors are equally likely when a worker is in a poor state. Violations and decision-making errors emerge as the most probable outcomes.
- Experience and Fitness Matter Most: Insufficient experience is highlighted as a particularly potent factor leading to unsafe behaviors, suggesting that simply having rules isn't enough if workers lack the practical know-how. Poor fitness for duty (encompassing competencies, experience, and even effects of medication or alcohol) is identified as the principal state that causes unsafe behaviors.
- BNs as a Powerful Diagnostic Tool: The study demonstrates the utility of Bayesian Networks in modeling complex, uncertain causal relationships in safety-critical domains, offering a more nuanced understanding than traditional statistical methods.
Key Results
- When a worker is in a poor state, the most vulnerable unsafe behavior is violation (probability not explicitly stated but ranked highest), followed by decision-making errors.
- Insufficient experience (IE) is identified as the most significant contributory factor to unsafe behavior.
- Poor fitness for duty (FD) is the principal worker state that causes unsafe behaviors.
- The study quantifies the impact of various worker states on unsafe behaviors. For instance, the average impact of inadequate safety awareness, mental fatigue, poor vigilance awareness, poor situational awareness, medical effect, illness, and insufficient experience on the four unsafe behaviors (skill-based errors, decision errors, perceptual errors, violations) are all substantial, ranging from approximately 0.50 to 0.525.
- When analyzing the four broader categories of worker states (AMS, APS, FD, PML), Fitness for Duty (FD) shows the highest impact across all unsafe behaviors, particularly decision errors.
Limitations and Future Work
The authors acknowledge a significant limitation: the "behavioural network" was built based on input from a limited number of experts due to the scarcity of comprehensive worker state data from accident investigations.
Future work suggested includes:
- Real-time Monitoring: Utilizing wearable devices (like Google Glasses or iWatches) to collect large-scale, real-time data on workers' health, habits, and states.
- Dynamic Bayesian Networks: Developing dynamic BNs to continuously evaluate risk levels of unsafe behaviors and propose timely interventions.
- Expanding Data Sources: Gathering more extensive data on worker states to build more robust and accurate models.
Summary updated