A Neural Network-Based Approach for Seismic-Induced Damage Assessment of Steel Liquid Storage Tanks

Abstract

Surrogate models, including neural network (NN), machine learning, and Kriging, are used in various fields to reduce the computational demand of risk assessment and uncertainty analysis. In civil engineering applications, surrogate models are usually trained on synthetic data generated with numerical simulation models, which might yield approximate responses and significant computational burdens. Post-disaster reconnaissance observations represent an alternative source of data that could be used to train a surrogate model without the need for numerical models. However, the limited number of reconnaissance observations available in the literature might yield challenges, such as unbalanced data distributions. This paper presents a surrogate-based approach for seismic-induced damage prediction, where postearthquake reconnaissance data are exploited to train a NN model. The approach is demonstrated on steel liquid storage tanks. Field data from past earthquake reconnaissance reports are first collected. Then, features representative of tank characteristics, seismic hazard parameters, and seismic-induced damage are extracted. A traditional two-layers NN model is built to map the relationship between tank characteristics, seismic hazard parameters, and seismic-induced damage. To tackle the challenge of unbalanced dataset, a cascade NN approach is proposed. In the cascade approach, two NN models are employed to predict the damage level of the steel tanks. The first NN returns a binary classification of the tank damage (i.e., damage, no damage). If the tank results damaged, a second NN identifies the level of damage (i.e., minor, severe, collapse). The performance of traditional and cascade NN approaches is compared using different metrics. Results demonstrate that the cascade NN strategy leads to more accurate damage predictions than the traditional NN approach.