Rearranged soft-introspective orthogonality regularized convolutional autoencoder for fault detection of multivariate processes

Modern industrial processes are developing towards scale and complexity, which poses high requirements for process monitoring. Autoencoder (AE) has gained increasing attentions as unsupervised deep learning models in recent years. However, those AE-based methods rarely considered influence of improp...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Advanced engineering informatics Ročník 69; s. 104026
Hlavní autoři: Wang, Chenmao, Lu, Zhiqiang, Yu, Jianbo
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Ltd 01.01.2026
Témata:
Convolutional autoencoder
Fault detection
Multivariate process
Orthogonality regularization
Multivariate process
Orthogonality regularization
Fault detection
Convolutional autoencoder
ISSN:1474-0346
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Modern industrial processes are developing towards scale and complexity, which poses high requirements for process monitoring. Autoencoder (AE) has gained increasing attentions as unsupervised deep learning models in recent years. However, those AE-based methods rarely considered influence of improper spatial order or view scale of process variables. Furthermore, they do not limit the information redundancy of latent features or consider the negative impact of training with normal data only. In this study, a new autoencoder model, i.e., rearranged soft-introspective orthogonality regularized convolutional autoencoder (ReSiOrCAE) is proposed to address the fault detection problem of large and complex multivariate processes. Firstly, an order rearrangement module is developed to optimize the spatial order of process variables. Secondly, a dual-path convolutional structure and an orthogonality regularization term are proposed to increase the details and diversity of extracted process features. A soft-introspective training strategy is used to enhance detection performance. The average fault detection rates (FDR) of ReSiOrCAE squared prediction error (SPE) and hotelling’s T2 statistics are 92.36 %, 99.60 %, 96.43 % and 91.50 %, 95.73 %, 95.00 % on the three benchmark processes, namely Tennessee Eastman process (TEP), continuous stirred tank reactor (CSTR) and hydraulic system, respectively, which shows the better performance compared to other typical models. The experiment results demonstrate the effectiveness of ReSiOrCAE in process fault detection.
ISSN:1474-0346
DOI:10.1016/j.aei.2025.104026