Novel Deep Learning-Based Method for Seismic-Induced Damage Detection

Abstract

In recent decades, the challenges of traditional visual inspection methods after catastrophic events, which are time- and money-consuming, have necessitated innovative approaches. As a result, a seismic-induced damage detection method utilizing deep learning has been developed to overcome the limitations of conventional techniques. Structure health monitoring (SHM) has emerged to address the limitations of the traditional methods of visual inspections, and among the most effective automatic feature extractor methods is Deep Learning Neural Networks (DLNNs). The DLNN method has proven highly effective compared to other methods, such as traditional methods used in damage detection when used as a feature extractor for seismic-induced damage detection. This study proposes a novel deep learning-based damage detection method for automatically extracting damage features from time series data, eliminating the need for intermediate preprocessing tools. The CNNs algorithm attains a validation accuracy of 91% when applied to a 7-story frame structure by subjecting the structures to different sets of incremental dynamic loading. The study investigates real-time applications, including environmental variables such as noise and temperature effects, examining unseen datasets of different earthquake groups and validating multiple structures in synthesis datasets. The algorithm is further investigated using the IASC-ASCE Benchmark experimental dataset conducted at the University of British Columbia laboratory. A comparative analysis is also performed in terms of time and performance on different deep learning algorithms, such as LSTM, 1D CNN, 2D CNNs and DNNs, while the 1D-CNNs showed the best performance. The results reveal that the proposed method effectively quantifies damage in different structures, including 7-story story steel and concrete structures, and the IASC-ASCE Benchmark dataset, with 93% validation accuracy. The study investigates different earthquake characteristics that affect deep learning performance, such as earthquake time step, and duration, while a specific group was examined to strengthen the claim and show 94% validation accuracy.