Unsupervised Damage Localization Using Autoencoders with Time-Series Data

In this paper, a non-linear autoencoder trained with time-series data for unsupervised damage localization based on residuals is investigated. Due to their sensitivity regarding small changes in the time-series, autoencoder offer a powerful tool for damage detection in Structural Health Monitoring (SHM). When it comes to output-only and unsupervised SHM, data-driven models struggle to properly localize the position of small damages. In an attempt to overcome these limitations, this study is performed using selected measurement data of a lattice tower called Leibniz University Structure for Monitoring (LUMO) under ambient excitation. Considering only data sets for similar material temperatures and wind speeds, the dependencies on environmental conditions are negligible. The model is trained using acceleration time-series as the input. As an extension of the model, the residuals are evaluated using the covariance. For each input signal and each residual time-series the covariance between them is calculated. The linear correlation of the input data to the residual increases the most for the sensor closest to the structural change. It can be shown that the data-driven model is able to locate all induced damages. The study presents a novel unsupervised data-driven damage localization technique using autoencoder with time-series data and correlations of the residual to the input data. This allows a localization of damage even when the manifestation of damage is not available.