Employment of bayesian networks for risk assessment of excavation processes in dense urban areas

The prediction of induced ground movements due to supported excavations are usually estimated by simulation analysis which relay upon a preliminary characterization of the soil mechanics. While empirical and semi – empirical methods are able to assess the excavation risk at the design stage, they cannot account for site – specific conditions and for information that become available as the excavation goes on. These latter capabilities can be very valuable if the digging site is located in dense urban areas as long as high ground movements can induce permanent damages in the surrounding buildings. Bayesian Networks have been profitably employed in order to update the reliability of geotechnical sites using measurement information [1]. The particular flexibility of Bayesian Networks is capable to employ Finite Element analysis and to account for spatial variability of the soil parameters [2]; also, this approach is able to account for simultaneous use of different types of evidence observations [3]. Aim of this work is to present a risk assessment strategy for multi stages excavation sites located in dense urban areas based on land survey measurements. The employment of laser scans and total stations makes possible to monitor the induced 3D displacements over a large area; using such kind of data for identification purposes leads to the updating of the excavation model at each digging stage. The updated geotechnical model is capable to check if the preliminary model is consistent with the evolution in time of the real geotechnical responses. Qualitative analysis of the updated parameters can be carried out in order to locate potential lacks of accuracy such as geometrical discrepancies or incongruous results of geotechnical tests. Most of all, the probabilistic characterization of the updated model allows a risk assessment: a simple reliability analysis evaluates the probability of occurrence of induced permanent damage in the surrounding buildings. In this sense, the Bayesian updating of the geotechnical model is able to give confidence about the excavation safety and to highlight possible critical issues. The Bayesian network is defined by means of a set of random variables. The weight of each surrounding building will be defined by theoretical probability distributions. Their calibration will be made by statistical investigation of historical data. The soil has been characterized by a geotechnical test campaign. The available data define the soil stratigraphy and the mechanical parameters of each layer. These characteristic values will be conveniently preprocessed in order to calibrate theoretical models of the soil probability distributions. The model response will be defined in terms of the vertical displacement of monitored points. Their a priori probability distributions will be evaluated by a Monte Carlo simulation. For this purpose, a non-linear finite element model will be employed. Thus, the excavation process will be simulated through OpenSees, a framework for finite element analysis purposes. Its features make possible to employ a very efficient pressure-dependent soil-type material and to simulate its interaction with the supporting structures [4]. Although, it is necessary to overcome some drawbacks about the network complexity and the linking procedure between the model and the recorded evidences, in order to perform an effective Bayesian updating. The network complexity is due to the topographical features of dense urban areas. Such kind of locations usually affects the model with a large number of random variables which could compromise its efficiency. This is due to the presence of several external loads and also the soil stratigraphy can entail a large amount of random variables. For this purpose, auxiliary nodes will be introduced into the network. First, the external loads will be modeled by means of global resultants; then, the soil parameters will concur in defining an equivalent soil column. Their aim is to limit the number of parents for the response nodes: in this sense, the network efficiency improves even with a larger set of variables as long as the parenthood relationships provide a nearly – chained network configuration. A further issue is about the evidences’ definition. Land survey devices usually provide a large amount of detected data which have to be statistically analyzed; thus, the identified displacements are not deterministic. In such cases, it is appropriate to employ probability distribution rather than deterministic values as provided information. For this reason, the network will be conveniently modified so that soft evidences can be properly handled [7]. The proposed procedure will be tested on a real case. The comparison between the results provided by the Bayesian network and the information gathered during the excavation process will show the strength of the procedure and its consistency. Moreover, a sensitivity analysis about the soft evidences will provide information