Real-time structural damage assessment using LSTM networks: regression and classification approaches

Structural health monitoring(SHM) techniques rarely consider the effect of ambient temperature, even though its impact on the structures being substantial. Moreover, typical modal or time-domain SHM approaches may delay the detection of damages endangering human lives due to their requirement of response time histories of sufficient length. Targeting prompt detection of structural anomalies, this article proposes a Long-Short-Term-Memory (LSTM)-based real-time approach that employs unsupervised LSTM prediction network for detection, followed by a supervised classifier network for localization. The prediction network is trained for one-step-ahead response prediction under ambient temperature conditions, and a novelty measure is devised using the usual prediction error threshold. Subsequently, damage is alarmed on encountering significant departure beyond this threshold. The damage is further localized with the classifier network. The approach is tested on a real bridge subjected to substantial thermal variation and the performance has been observed to be prompt and reliable under different operating conditions.