Encoding Time-Series Ground Motions as Images for Convolutional Neural Networks-Based Seismic Damage Evaluation

Abstract

Traditional methods for seismic damage evaluation require manual extractions of intensity measures (IMs) to properly represent the record-to-record variation of ground motions. Contemporary methods such as convolutional neural networks (CNNs) for time series classification and seismic damage evaluation face a challenge in training due to a huge task of ground-motion image encoding. Presently, no consensus has been reached on the understanding of the most suitable encoding technique and image size (width × height × channel) for CNN-based seismic damage evaluation. In this study, we propose and develop a new image encoding technique based on time-series segmentation (TS) to transform acceleration (A), velocity (V), and displacement (D) ground motion records into a three-channel AVD image of the ground motion event with a pre-defined size of width × height. The proposed TS technique is compared with two time-series image encoding techniques, namely recurrence plot (RP) and wavelet transform (WT). The CNN trained through the TS technique is also compared with the IM-based machine learning approach. The CNN-based feature extraction has comparable classification performance to the IM-based approach. WT 1,000 × 100 results in the highest 79.5% accuracy in classification while TS 100 × 100 with a classification accuracy of 76.8% is most computationally efficient. Both the WT 1,000 × 100 and TS 100 × 100 three-channel AVD image encoding methods are promising for future studies of CNN-based seismic damage evaluation.