A 1D-CNN deep learning framework for seismic collapse prediction of jacket offshore platforms with Bayesian neural architecture search

This paper aims to develop and evaluate the one dimensional convolutional neural network (1D-CNN) deep learning (DL) model for earthquake accelerogram time series classification, with the ultimate goal of predicting seismic limit state exceedance in Jacket Type Offshore Platforms (JTOP). The application of earthquake time series as the DL model input feature is not common in the fast seismic performance evaluation of structures. Additionally, in practical scenarios, the predictive performance of 1D-CNN is heavily reliant on the selection of appropriate model architecture and hyperparameters. Therefore, this study proposes a framework for systematically searching the neural architecture and hyperparameters of deep learning models, particularly 1D-CNN, through Bayesian optimization. Since manual search for model components can be a difficult task that may require numerous trial-and-error attempts or parametric studies, the aim of this framework is to streamline and organize the typical workflow of developing DL models to ensure optimal performance. Moreover, the comprehensive nested cross-validation method has been utilized to achieve an unbiased evaluation. The focus of this study is on the seismic collapse prediction of Jacket Type Offshore Platforms with 1D-CNN, and to fulfill this objective, a stochastic optimization algorithm is proposed to perform the stratified K-fold cross validation, specifically developed for the dataset preparation based on the incremental dynamic analysis (IDA). For comparison purposes, the presented framework is also applied for the hyperparameter tuning of Multilayer Perceptron (MLP) and Support Vector Machine (SVM) techniques. The effectiveness of the presented framework is evaluated for a case study JTOP structure subjected to an ensemble of earthquake records. The results demonstrate the capability of 1D-CNN model, developed using the proposed framework, for computing the collapse fragility curves of JTOPs.