Deep Transfer Learning and Time-Frequency Characteristics-Based Identification Method for Structural Seismic Response

Abstract

The cost of dedicated sensors has hampered the collection of the high-quality seismic response data required for real-time health monitoring and damage assessment. The emergence of crowdsensing technology, where a large number of mobile devices collectively share data and extract information of common interest, may help remove such obstacles and mitigate the seismic hazard. The present study proposes a crowdsensing-oriented vibration acquisition and identification method based on time–frequency characteristics and deep transfer learning. It can distinguish the responses during an earthquake event from vibration under serviceability conditions. The core classification process is performed using a combination of wavelet transforms and deep transfer networks. The latter were pre-trained using finite element models calibrated with the monitored seismic responses of the structures. The validation study confirmed the superior identification accuracy of the proposed method.