PCA-TANN with model-based transfer learning for predicting blast load time series on structures

Abstract

A comprehensive understanding of blast load information is crucial for evaluating structural response. Machine learning has demonstrated its efficiency in predicting blast loads. However, the high costs associated with acquiring blast load data limit the development of machine learning. A novel method, PCA-TANN, is proposed for predicting blast load time series on structures and improving the problem of insufficient training data by integrating transfer learning concepts and combining artificial neural networks (ANN) with principal component analysis (PCA). The main idea of this model involves preprocessing the time series data using PCA to extract essential features, employing artificial neural networks based on transfer learning (TANN) for knowledge transfer from the previous task to the new task, and integrating a few preprocessed data of the new task for training. The blast load data obtained from a single square column is utilized to assess the performance of PCA-TANN and PCA-ANN models. The findings demonstrate that in situations with limited data availability, PCA-TANN reasonably replicates the blast load characteristic, resulting in accurate time series predictions of overpressure with R2 scores exceeding 0.85. This performance significantly surpasses that of the used PCA-ANN model, which achieves R2 scores below 0.4.