A Bayesian Network Approach for Probabilistic Safety Analysis of Traffic networks
This thesis consists mainly of two parts. The first one is based on a review of concepts and models that have been useful to carry out this study. The second is the one that covers the majority of the work, where a new model of probabilistic analysis for the study of road safety based on Bayesian Ne...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2017 |
| País: | España |
| Institución: | Universidad de Cantabria (UC) |
| Repositorio: | UCrea Repositorio Abierto de la Universidad de Cantabria |
| Idioma: | inglés |
| OAI Identifier: | oai:repositorio.unican.es:10902/13112 |
| Acceso en línea: | http://hdl.handle.net/10902/13112 |
| Access Level: | acceso abierto |
| Palabra clave: | Seguridad de carreteras Redes bayesianas Seguridad vial Análisis probabilístico de seguridad Road safety Bayesian networks Probabilistic safety analysis |
| Sumario: | This thesis consists mainly of two parts. The first one is based on a review of concepts and models that have been useful to carry out this study. The second is the one that covers the majority of the work, where a new model of probabilistic analysis for the study of road safety based on Bayesian Networks is presented. Bayesian networks use the probabilistic structure of a multidimensional random variable based on an acyclic graph and a set of conditional probabilities to perform a probabilistic safety analysis. Due to the great advantages that they present with respect to other methods used, such as regression or failure tree methods, they have been used in the last years in many different fields: artificial intelligence, tunneling processes, biomedicine, nuclear plants, and more recently on railway lines. What is intended in this thesis is to perform a safety analysis on roads so this first part focuses on models of Bayesian networks applied to road safety. In this way, models with different purposes are presented: a) predicting the frequency with which accidents of different types occur, b) classifying traffic accidents according to their severity, c) analyzing and preventing accidents, and d) safety. On the other hand, without the knowledge of certain aspects on graphs would not be possible a correct construction of a model of Bayesian networks. For this reason and thus to be able to clarify concepts and methods used, different explanatory illustrations are presented in which variables used later in the proposed model are utilized. The second part presents a new model for the study of road safety using a probabilistic analysis based on Bayesian networks. The problem of the safety analysis is an undoubtedly random problem, since practically all the variables that intervene in the same one are random. This requires evaluating probabilities of occurrence of events and frequencies associated with different intervening elements. A first problem arises when representing the dependencies between the variables. Fault trees have some important limitations, including failure to easily represent common causes of failure. The tree structure, i.e. with open branches, does not allow closing them, which would be necessary to reproduce the common causes without replicating the corresponding variables. In contrast, Bayesian networks do not have this limitation and allow their reproduction without the need to replicate these variables. Another notable advantage is that, when using directed graphs, they can be closed and the joint probability of all variables determined by the conditional probabilities of each node given by their parents. In addition, these networks can reproduce any dependency structure without producing incompatibilities, which can occur when the definition of joint probability is made with arbitrary conditional distributions. Another great advantage is that all of the above conditional probabilities can be defined independently. Finally, it should be noted that there are very powerful and even free software packages, which have already been tried and tested, and which allow the implementation of these computer structures without any extra effort. All these advantages are what have led to choosing the Bayesian networks as the optimal model to solve this problem. The Bayesian network model reproduces not only all the existing elements on the road but the driver's behaviour when he is driving through it. Each element contributes a set of variables according to their type. For example, curves include their radius, length, direction, etc., the signals include their state, the driver's decisions upon seeing them, the associated speeds, the distances between signals, and so on. Due to the great importance of human error in the field of safety, modeling variables associated with driver's behaviour are introduced, such as driver's tiredness and attention; as well as the type of driver or decision of the adopted speed or the presence of a signal. All variables are considered as random and their dependencies are reproduced by the Bayesian network, so that any set of probabilities can be calculated through a process of forward marginalization. The sets of conditional probabilities of the variables, given their parents, are established by means of closed formulas that allow to quantify the Bayesian network. To reduce the complexity of the problem, we propose to use a method that divides the Bayesian network into small parts, such that the complexity of the problem becomes linear in the number of elements. This is crucial to deal with real cases where the number of variables can be measured in thousands. The probability of incidents related to the different road sections is calculated according to an equivalent number of severe incidents, so that the most critical elements can be identified and ranked in order of importance. This allows to obtain very relevant information to improve the safety and to save time and money in the measures that are necessary to adopt to improve certain roads. In addition, when an accident occurs, the Bayesian network can help identify its causes through a process of inference propagation backwards. Different examples of Spanish roads, A-67, N-611 and CA-182, are represented to expose the operation of the model that arises. In particular, a detailed study of the regional network in Cantabria, CA-131, CA-132 and CA-142, is carried out by means of this new method. |
|---|