Structural damage detection with autoencoding neural networks
Palavras-chave:
sparse autoencoder, structural health monitoring, damage detectionResumo
Structural Health Monitoring (SHM) is a growing field in civil engineering, having relevance for the de-
tection of changes in the state of structures, including the identification of possible damage conditions. Commonly
used SHM strategies involve employing Artificial Intelligence (AI) techniques on raw dynamic data measured
from structures to perform classifications or extract features from the original data. Among the AI algorithms for
SHM, autoencoding neural networks, or simply autoencoders, have been identified as promising solutions, being
the focus of this article. An autoencoder is designed to reproduce its inputs as closely as possible after unsuper-
vised training. This characteristic is a useful tool to extract features that represent the original data with lower
dimensionality, which facilitates classifications through statistical methods. A sparse autoencoder (SAE) is a type
of autoencoder that includes a sparsity penalty at its training process. In that way, this paper presents a method-
ology to detect structural damage by using sparse autoencoders as parameter extractors applied to signals in the
frequency domain, combined with the Mahalanobis distance to perform an unsupervised classification. Tests per-
formed with data extracted from the Z24 bridge have shown promising results in detecting changes in the structural
states, demonstrating the potential of SAE for SHM systems.