A Novel Machine Learning Based Framework for the Seismic Risk Assessment of Industrial Plant

Abstract

Abstract In last decades, Machine learning techniques (ML) have been extensively studied as a possible reliable family of regression methods. Recently ML, which originated in the field of computer science, has been widely investigated as an alternative and helpful method for solving engineering problems. The main feature of ML techniques is learning from a dataset, containing a certain amount of input/output fidelity data, in order to predict future output data related to a new set of input data in cost-effective computation time. With these premises, the application of this type of techniques to the civil engineering field, where a sufficient amount of data is available by means numerical and experimental simulations and from structural monitoring, seems interesting. Along this vein, the present paper aims to investigate the suitability of a novel ML based framework for the risk assessment of a Non Structural Component (NSC) of an industrial plant. The proposed method, which aims to solve some critical issues associated to the traditional risk assessment methodology, has the follow main advantages: (i) reduction of computational time by means an Artificial Neural Network (ANN) surrogate model; (ii) avoidance of prior assumptions on the distribution of fragility curves, sampling a large amount of data from the ANN surrogate model; (iii) avoidance of the record to record variability of the seismic input through a new record selection algorithm; (iv) integration of hazard and vulnerability analysis in the same framework. In this respect, a comprehensive sensitivity analysis of the ANN input parameters is performed (features selection), identifying the type and number of Intensity Measures (IMs) that represent the seismic input best related to the structural output, using Principal Component Analysis (PCA). Furthermore, a multiple stripe analysis is performed on a nonlinear finite element model (FEM) of the structure, deriving the set of data used to train and validate the ANN. Then, two different surrogate models in series are derived, investigating the architecture of the models (i.e. the number of hidden layers and parameters weight), to take into account both the dependence between the selected IMs and the relation between the IMs and the structural response. Different training and test subsets are used to derive the surrogate models, to find the best-performing structure. A sampling of the seismic input parameters of the ANN is obtained followed their probability distribution by means Metropolis-Hastings algorithm, to take in account the hazard of the site, and the risk assessment is directly carried out from the observed damage evaluated. The proposed framework, which represents an interesting alternative for seismic risk assessment, is finally applied to an industrial NSC.