An Investigation of the Effect of Measurement Interval on the Autoencoder Based Damage Detection in Uncontrolled Structural Health Monitoring

Abstract

Unsupervised damage detection in uncontrolled, outdoor environmental and operational conditions (EOCs) is crucial for practical structural health monitoring. While previous research has explored autoencoder-based unsupervised damage detection methods, they require training data only from pristine conditions. In long-term monitoring, irregular environmental and operational conditions, as well as variations in damage, may make it difficult to satisfy this requirement. In this paper, we propose a novel autoencoderbased approach that uses training data containing regular and irregular environmental and operational conditions, as well as damage variations. We also investigate the impact of various factors, such as training epoch, damage duration, and measurement interval on the accuracy of damage detection. Our results indicate that our proposed framework achieves an AUC score of over 0.95 when the measurement interval is around 860 seconds per measurement. Interestingly, this score decreases both when we sampling faster and slower.