TL;DR
This paper introduces a novel hybrid framework combining an improved Human Factors Analysis and Classification System (HFACS) with Bayesian Networks to analyze the complex causal chains of waterborne traffic accidents. By applying this data-driven approach to real-world accident data, the study identifies key contributing factors and their interdependencies, offering a more nuanced understanding for enhanced safety measures.
What Problem Does This Paper Address?
Waterborne transportation is crucial for global trade, but accidents pose significant risks to life, property, and the environment. These accidents are rarely caused by a single factor; instead, they arise from a complex interplay of human, organizational, and environmental elements. Existing research often focuses on single accident types or regions, or uses singular analytical frameworks, limiting a comprehensive understanding of the intricate causal chains. This paper aims to bridge this gap by developing a more generalized, multi-dimensional analytical framework to uncover these complex relationships and inform accident prevention strategies.
Methods
The core of this research is a hybrid framework that integrates two key methodologies:
- Improved HFACS (Maritime Traffic Accident Analysis and Classification System - MTAACS): The authors adapt the standard HFACS model, which is widely used in aviation safety, to better suit the unique characteristics of waterborne accidents. Key improvements include explicitly incorporating an "accident level" for classification and considering external environmental influences at the "preconditions for unsafe behavior" stage. The MTAACS categorizes causal factors into five hierarchical levels: organizational influences, inadequate supervision, preconditions for unsafe behaviors, unsafe behaviors, and the accident layer. Importantly, it models relationships between adjacent levels as nonlinear, acknowledging the complex interactions.
- Bayesian Networks (BN): This probabilistic graphical model is used to represent and analyze the dependencies between the identified causal factors. The BN structure is learned from data using statistical tests (specifically, the chi-square test for structural learning), and then parameters are estimated using Maximum Likelihood Estimation. This allows for the quantification of relationships and the identification of causal pathways.
The overall process involves initializing an accident dataset, calculating factor occurrences, performing correlation analysis between factor pairs, constructing and applying the Bayesian network, and finally outputting the analysis results.
Datasets
The study utilizes a dataset of 756 investigative reports on water traffic accidents along China's coast, collected from the official website of the China Maritime Safety Administration between June 2015 and February 2023. Strandings and contact accidents were excluded due to insufficient sample data. For collision accidents, each vessel was treated as an independent research sample.
Key Insights and Ideas
- Non-linear Causality is Key: The paper emphasizes that waterborne accidents are not linear sequences of events. The improved HFACS (MTAACS) explicitly models these nonlinear relationships, allowing for a more realistic representation of how multiple factors can interact and influence outcomes.
- Hybrid Approach Unlocks Deeper Insights: Combining the structured categorization of HFACS with the probabilistic reasoning of Bayesian Networks provides a powerful tool. HFACS helps identify and classify potential causes, while BN quantifies their interdependencies and probabilities, leading to a more comprehensive understanding than either method alone.
- Context Matters: The adaptation of HFACS to MTAACS highlights the need for domain-specific adjustments in analytical frameworks. Explicitly including an "accident level" and considering environmental factors are crucial for waterborne accident analysis.
- Data-Driven Validation: The reliance on a large dataset of real accident reports lends significant credibility to the findings, moving beyond theoretical models to empirically grounded conclusions.
Key Results
The study identified a total of 38 causal factors categorized into the five levels of MTAACS. The analysis revealed:
- Accident Type Distribution: Collision accidents, "other accidents," and sinking accidents were found to be the most frequent.
- Key Causal Factors: At the organizational level, "inadequate safety management execution" (A5), "improper staffing" (A1), and "inadequate education and training" (A2) were identified as critical. For unsafe behaviors, "improper lookout" (D1), "misjudgment of hazards" (D3), and "failure to take early action" (D9) were most prominent.
- Sensitivity Analysis: For collision accidents, "insufficient safety awareness" (C2) showed the highest sensitivity. For contact accidents, "improper collision avoidance behavior" (D4) was most influential. For fire/explosion accidents, "improper use of equipment" (D7) was key. For windstorm accidents, "inadequate typhoon prevention measures" (D12) were most critical. For sinking accidents, "improper emergency measures" (D11) were paramount.
- Causal Chains: The study identified varying numbers of causal chains for different accident types: 12 for collisions, 3 for contact, 6 for fire/explosions, 2 for windstorms, 4 for sinking, and 7 for other accidents.
Limitations and Future Work
- Data Specificity: While the dataset is substantial, it is drawn from China's coastal waters. The authors suggest future research should include comparative analysis with other maritime regions to explore similarities and differences in causal chains across diverse environments.
- Methodological Advancements: The paper proposes leveraging advanced technologies like artificial intelligence and machine learning for future analyses. This could further deepen the understanding of causal chains and potentially lead to more sophisticated accident prediction and prevention systems, especially with the rise of Maritime Autonomous Ships (MASSs).
- Data Availability: The authors note that the data presented in the study is available on request from the corresponding author.
Summary updated