Uncertainty-aware structural damage warning system using deep variational composite neural networks

Structural health monitoring (SHM) is, without a doubt, one of the most important assets for building resilient communities. The vast and rapidly advancing research in data science and machine learning has provided researchers in the civil engineering community with various tools that can facilitate the processing of significant amounts of gathered data. However, deep learning models are prone to mistakes, and with the catastrophic consequences that can happen due to damage misidentification, damage diagnosis models' predictions should not be taken for granted. In this study, we present an uncertainty-aware early-warning system that can provide near real-time SHM. The system utilizes a deep composite encoder-decoder network that combines elements from convolutional neural networks, recurrent neural networks, and variational inference (VI) to provide damage index distributions. The framework can detect anomalies in the structural system during seismic events and provide a measure of uncertainty that can be used to question the model's predictions. To assess the system's validity and practicality, we apply our proposal to three real structures, two of which suffered damage during the 1994 Northridge earthquake. We found that the early warning system delivers an accurate, yet cautious, continuous monitoring that is capable of sending warning signals when damage occurs in the course of seismic events. Source code is available at: https://github.com/keltouny/VSCAN.