Fusing convolutional generative adversarial encoders for 3D printer fault detection with only normal condition signals

Collecting data from mechanical systems in abnormal conditions is expensive and time consuming. Consequently, fault detection approaches based on classical supervised learning working with both normal and abnormal data are not applicable in some conditionbased maintenance tasks. To address this prob...

Full description

Bibliographic Details
Authors: Li, Chuan, Cabrera, Diego, Sancho Caparrini, Fernando, Sánchez, René-Vinicio, Cerrada, Mariela, Long, Jianyu, Oliveira, José Valente de
Format: article
Status:Versión enviada para evaluación y publicación
Publication Date:2021
Country:España
Institution:Universidad de Sevilla (US)
Repository:idUS. Depósito de Investigación de la Universidad de Sevilla
OAI Identifier:oai:idus.us.es:11441/107188
Online Access:https://hdl.handle.net/11441/107188
https://doi.org/10.1016/j.ymssp.2020.107108
Access Level:Open access
Keyword:fault Detection
3D printer
Condition-based maintenance
Convolutional Neural Networks (CNN)
Adversarial learning
Description
Summary:Collecting data from mechanical systems in abnormal conditions is expensive and time consuming. Consequently, fault detection approaches based on classical supervised learning working with both normal and abnormal data are not applicable in some conditionbased maintenance tasks. To address this problem, this paper proposes Fusing Convolutional Generative Adversarial Encoders (fCGAE) method to create fault detection models from only normal data. Firstly, to obtain an adequate deep feature space, encoder models based on 1D convolutional neural networks are created. Then, these encoders are optimized in an unsupervised way through Bidirectional Generative Adversarial Networks. Finally, the multi-channel features collected from the system are merged with One-Class Support Vector Machine. fCGAE is applied to fault detection in 3D printers, where experimental results in two fault detection cases show excellent generalization capabilities and better performance compared to peer methods.