تجزیه و تحلیل ریسک در طی ساخت تونل با استفاده از شبکه های بیزی: مطالعه موردی مترو پورتو

This paper presents a methodology to systematically assess and manage the risks associated with tunnel construction. The methodology consists of combining a geologic prediction model that allows one to predict geology ahead of the tunnel construction, with a construction strategy decision model that allows one to choose amongst different construction strategies the one that leads to minimum risk. This model used tunnel boring machine performance data to relate to and predict geology. Both models are based on Bayesian Networks because of their ability to combine domain knowledge with data, encode dependencies among variables, and their ability to learn causal relationships. The combined geologic prediction–construction strategy decision model was applied to a case, the Porto Metro, in Portugal. The results of the geologic prediction model were in good agreement with the observed geology, and the results of the construction strategy decision support model were in good agreement with the construction methods used. Very significant is the ability of the model to predict changes in geology and consequently required changes in construction strategy. This risk assessment methodology provides a powerful tool with which planners and engineers can systematically assess and mitigate the inherent risks associated with tunnel construction.

There is an intrinsic risk associated with tunnel construction because of the limited a priori knowledge of the existing subsurface conditions. Although the majority of tunnel construction projects have been completed safely there have been several incidents in various tunneling projects that have resulted in delays, cost overruns, and in a few cases more significant consequences such as injury and loss of life. It is therefore important to systematically assess and manage the risks associated with tunnel construction. A detailed database of accidents that occurred during tunnel construction was created by Sousa (2010). The database contains 204 cases all around the world with different construction methods and different types of accidents. The accident cases were obtained from the technical literature, newspapers and correspondence with experts in the tunneling domain. Knowledge representation systems (or knowledge based systems) and decision analysis techniques were both developed to facilitate and improve the decision making process. Knowledge representation systems use various computational techniques of AI (artificial intelligence) for representation of human knowledge and inference. Decision analysis uses decision theory principles supplemented by judgment psychology (Henrion, 1991). Both emerged from research done in the 1940s regarding development of techniques for problem solving and decision making. John von Neumann and Oscar Morgensten, who introduced game theory in “Games and Economic Behavior” (1944), had a tremendous impact on research in decision theory. Although the two fields have common roots, since then they have taken different paths. More recently there has been a resurgence of interest by many AI researchers in the application of probability theory, decision theory and analysis to several problems in AI, resulting in the development of Bayesian Networks and influence diagrams, an extension of Bayesian Networks designed to include decision variables and utilities. The 1960s saw the emergence of decision analysis with the use of subjective expected utility and Bayesian statistics. Howard Raiffa, Robert Schlaifer, and John Pratt at Harvard, and Ronald Howard at Stanford emerged as leaders in these areas. For instance Raiffa and Schlaifer’s Applied Statistical Decision Theory (1961) provided a detailed mathematical treatment of decision analysis focusing primarily on Bayesian statistical models. Pratt et al. (1964) developed basic decision analysis. while Eskesen et al. (2004) and Hartford and Baecher (2004) provide good summaries on the different techniques (fault trees, decision trees, etc.) that can be used to assess and manage risk in tunneling. Various commercial and research software for risk analysis during tunnel construction have been developed over the years, the most important of which is the DAT (Decision Aids for Tunneling), developed at MIT in collaboration with EPFL (Ecole Polytechnique Fédérale de Lausanne). The DAT are based on an interactive program that uses probabilistic modeling of the construction process to analyze the effects of geotechnical uncertainties and construction uncertainties on construction costs and time. (Dudt et al., 2000 and Einstein, 2002) However, the majority of existing risk analysis systems, including the DAT, deal only with the effects of random (“common”) geological and construction uncertainties on time and cost of construction. There are other sources of risks, not considered in these systems, which are related to specific geotechnical scenarios that can have substantial consequences on the tunnel process, even if their probability of occurrence is low. This paper attempts to address the issue of specific geotechnical risk by first developing a methodology that allows one to identify major sources of geotechnical risks, even those with low probability, in the context of a particular project and then performing a quantitative risk analysis to identify the “optimal” construction strategies, where “optimal” refers to minimum risk. For that purpose a decision support system framework for determining the “optimal” (minimum risk) construction method for a given tunnel alignment was developed. The decision support system consists of two models: a geologic prediction model, and a construction strategy decision model. Both models are based on the Bayesian Network technique, and when combined allow one to determine the ‘optimal’ tunnel construction strategies. The decision model contains an updating component, by including information from the excavated tunnel sections. This system was implemented in a real tunnel project, the Porto Metro in Portugal.

A decision support framework for assessing and avoiding risks in tunnel construction was developed and successfully applied in a case study. The decision support framework consists of the geology prediction model and the construction strategy decision model both of which are based on Bayesian Networks. TBM performance data are used to predict geology, which is then used to help decide on the construction method involving the lowest risk. The presented risk model contains two models, a geological prediction model and a decision model. The geological model was trained (or calibrated) with the data from a specific project (the Porto Metro). Afterwards, the models were applied, i.e. tested on another section of the Porto Metro. The data that were used to test the model were not those used to the train the model. The results of the predictions of geology on the part of tunnel that was not used to train the model can predict changes in geology. The application to the Porto Metro tunnel case, in which several accidents occurred, shows that the decision support framework fulfills its objectives. Specifically the results show that the model can predict changes in geology and that it suggests changes in construction strategy. This is most visible in the zone of accidents 2 and 3, where the model accurately predicts the change in geology and occurrence of soil. The “optimal” construction strategy determined by the combined risk assessment model is EPBM in closed mode, i.e. with a fully pressurized face, in the areas where accident 2 and 3 occurred, and not what was actually used during construction, EPBM in open/semi-closed mode. This difference is due to the fact that during the actual construction there was no effective system to predict changes in geology and therefore adapt the construction strategy. Clearly the question arises how the proposed methodology process would work in other cases. If the geological prediction model were applied elsewhere (in another type of geology) it would need calibration. There is also the issue of not having data to calibrate the model at the beginning of construction. A way to use and calibrate this type of prediction model in cases where initial data do not exist (from a nearby project or similar geology) would be to use at the beginning of the construction subjective probabilities given by the experts (e.g. What is the probability that one is in geology G1, if the penetration rate is high (i.e. greater than a certain value), etc.), and then update these probabilities as the construction progresses.