Deep reference autoencoder convolutional neural network for damage identification in parallel steel wire cables

This paper addresses the challenge of overcoming the limited availability of training samples for the damage state of parallel steel wire cables in practical engineering and the difficulty of detecting minor damages. To tackle this issue, we propose an unsupervised deep learning damage identificatio...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Structures (Oxford) Ročník 57; s. 105316
Hlavní autoři: Xue, Songling, Sun, Yidan, Su, Teng, Zhao, Xiaoqing
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Ltd 01.11.2023
Témata:
Damage identification
DRACNN
Parallel steel wire cable
Unsupervised deep learning
Unsupervised deep learning
Parallel steel wire cable
Damage identification
DRACNN
ISSN:2352-0124, 2352-0124
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:This paper addresses the challenge of overcoming the limited availability of training samples for the damage state of parallel steel wire cables in practical engineering and the difficulty of detecting minor damages. To tackle this issue, we propose an unsupervised deep learning damage identification technique called Deep Reference Autoencoder Convolutional Neural Network (DRACNN) for analyzing the damage state of parallel steel wire cables in bridge engineering. The DRACNN method utilizes multi-dimensional cross-correlation function (CCF) derived from acceleration signals at various health stages as input to train the network structure and obtain optimal parameters. Subsequently, we analyze the layer decomposition to identify neurons in the lowest hidden layer indicating damage. The neuronal change information is then extracted using an Exponentially Weighted Moving Average (EWMA) Control Chart to determine the damage state of the structure. Finally, we present a comprehensive numerical analysis describing the method's flow and network architecture and demonstrate the feasibility of this approach through experiments.
ISSN:2352-0124
2352-0124
DOI:10.1016/j.istruc.2023.105316