مدل سازی شبکه های بیزی از یافته های بازرسی کنترل بندر ایالتی و دخالت حوادث کشتی

The paper utilizes Port State Control inspection data for discovering interactions between the numbers of various types of deficiencies found on ships and between the deficiencies and ship’s involvement in maritime traffic accidents and incidents. Bayesian network models for describing the dependencies of the inspection results, ship age, type, flag, accident involvement, and incidents reported by the Vessel Traffic Service are learned from the Finnish Port State Control data from 2009–2011, 2004–2010 Baltic Sea accident statistics and the reported Gulf of Finland Vessel Traffic Service incidents within 2004–2008. Two alternative Bayesian network algorithms are applied to the model construction. Further, additional models including a hidden variable which represents the complete system and its safety features and which links the accident and incident involvement and Port State Control findings are presented. Based on model-data fit comparisons and 10-fold cross-validation, a constraint-based learning algorithm NPC mainly outperforms the score-based algorithm repeated hill-climbing with random restarts. For the highest scoring models, mutual information and influence of evidence analyses are conducted in order to analyze which network variables and variable states are the most influential ones for determining the accident involvement. The analyses suggest that knowledge on ship type, the Port State Control inspection type and the number of structural conditions related deficiencies are among the ones providing the most information regarding accident involvement and the true state of the hidden system safety variable.

Maritime traffic accidents have fortunately been rather rare events. For example, approximately 0.7 accidents per 1000 port ship calls have taken place in the Gulf of Finland, the densely trafficked but shallow Baltic Sea area between Finland, Russia and Estonia (Kujala et al., 2009 and Kuronen et al., 2009). However, if occurring, a maritime traffic accident could have adverse outcomes resulting in the loss of lives or a severe damage to the environment (Helle, Lecklin, Jolma, & Kuikka, 2011). The occurrence of accidents in a system can be seen as an indicator of the system safety, or rather the absence of safety. However, past accident frequency alone is not a very informative indicator on how accidents occur or what makes a system safe. Maritime traffic system, or even a single ship, is a complex socio-technical system. Since Heinrich (1931) presented the Domino theory, various frameworks and models of how and why accidents occur in such a system have been proposed, e.g. Perrow, 1984, Reason, 1990, Rasmussen, 1997, Hollnagel, 1998, Dekker, 2002, Leveson, 2004 and Hollnagel, 2009; for a summary, see Qureshi (2007). The recent models approach the problem from a systemic viewpoint and state that the accidents are a result of variability in the system, a property which is inevitable, and even desired, as the variability also prevents many accidents (Hollnagel, 2004). Furthermore, according to the systemic accident models it is impossible and unadvisable to describe the cause-effect relationships of accidents within a complex system (Hollnagel, 2009 and Dekker, 2011). Instead of attempting to develop a causal model of a maritime accident occurrence, modeling any potential patterns between different system safety indicators, given that the indicators truly measure safety, could provide indirect albeit potentially useful information for the purposes of safety management. Further, as such a model contains factors describing measurable features, its validation might be less challenging. Still, one should consider the inevitable epistemic and aleatory uncertainty due to multiple reasons such as the chosen modeling approach itself, the necessary modeling assumptions and simplifications, incomplete knowledge of the system components and their dependencies and due to the system complexity and variation between the shipping companies, vessels, crew properties, and the dynamic circumstances, for instance. Also, some relevant factor might be unobservable and thus no data is available on that. An attractive quantitative modeling technique that can present relatively complex, not necessarily causal dependencies and cope with at least some of these uncertainties and unobservable variables while also having a qualitative, graphical dimension, is Bayesian belief networks (Pearl, 1988). Safety indicators are measurable features which can be utilized in describing the safety state of a system (e.g. Øien, Utne, & Herrera (2011)). If data is available, machine learning techniques could be employed to discover the dependencies between the relevant safety indicators – either with a purely data-driven perspective or by incorporating prior information using Bayesian techniques. Previous studies have applied machine learning to discovering various maritime safety-related information from data (Le Blanc and Rucks, 1996, Blanc et al., 2001, Grech et al., 2002, Tzannatos and Kokotos, 2009, Hashemi et al., 1995, Kokotos and Smirlis, 2005 and Kelangath et al., 2011; Antao, Guedes Suares, Grande, & Trucco, 2008; Mascaro, Nicholso, & Korb, 2013). However, although some initial work has been published on utilizing machine learning in maritime traffic accident modeling (Kristiansen, 2010), their applications to probabilistically describe the dependencies of various factors and maritime accidents without a causal interpretation remain absent. For maritime traffic safety controlling purposes, one set of potential safety indicators results from Port State Control inspections. Port State Control (PSC) inspection, conducted by the port state authority when foreign ships visit their port, checks the condition, the equipment, manning and the operation of foreign state vessels for verifying that the aforementioned aspects on board comply with international regulations (IMO, 2011). If the inspection reveals deficiencies that pose a safety hazard, the ship may be detained at the port until the deficiencies have been rectified. This paper continues the work conducted by the authors in Hänninen and Kujala (submitted for publication) and explores how Port State Control inspection findings on different areas of inspection are linked to each other and to maritime traffic accidents and incidents using Bayesian belief networks. Previously published studies utilizing PSC data (Mejia et al., 2010, Knapp and Franses, 2007, Li et al., 2010 and Soma, 2004) have not studied or modeled these connections in detail. Further, the paper introduces Bayesian network models where “accident causation mechanisms” or “system variability” is presented with one immeasurable so-called hidden variable which is linked to the safety indicators, i.e., the PSC deficiencies, incident and accident involvement. The resulting Bayesian network models can be used as an aid in safety management, for example when a certain number of PSC deficiencies of one – or multiple – type(s) have been observed and one wants to examine what this observation tells about the other deficiencies, how the proneness to accidents changes given various deficiency observations, or how the deficiencies and other model variables differ between ships which have been involved in accidents or incidents and the ones which have not. The rest of the paper is organized as follows. The data and the methods used in the study are described in Section 2. Section 3 presents the learned models. These results are further discussed in Section 4, followed by some conclusions and recommendations for the future research in Section 5.

This study has explored the dependencies of PSC inspection findings and ship’s involvement in accidents and incidents by learning Bayesian networks from the inspection, accident and incident data with two alternative learning algorithms. In addition to models which only include variables included in the input data, networks have been constructed with an additional hidden variable representing the ship system and its safety features, through which accident and incident involvement and the PSC deficiencies are connected. The resulted models can be used for examining e.g. how numbers of different types of deficiencies, detention from an inspection, being reported by a VTS officer, accident involvement and the ship age, type and flag are dependent and how observing the true states of network variables affect the probabilities of other variable states. It can be concluded that whereas the hidden variable models fit best to the utilized dataset, the Bayesian network without the hidden variable and learned with the NPC algorithm has the best generalizability. In general, the interactions between the network variables seem to be rather weak, which is not surprising given the high probabilities of zero deficiencies and no detention in the PSC data. Nevertheless, it should be noted that PSC inspection findings is only one set of maritime traffic safety related indicators. Further, its information mainly presents the conditions onboard and only indirectly contains information on safety on the higher levels of organization. Also, the inspection data is a result of the inspectors’ interpretations, which might be biased and thus it is uncertain whether the recorded deficiencies are truly indicators of safety - or the lack of it. Therefore, the work for developing a Bayesian network tool for maritime safety management will continue by utilizing additional ship safety and safety management -related information sources