High and Robust Fault Detection via Polynomial Approximated Isomap Embeddings
Computer Aided Chemical Engineering, Elsevier B.V., ss.607-612, 2023
- Yayın Türü: Kitapta Bölüm / Araştırma Kitabı
- Basım Tarihi: 2023
- Doi Numarası: 10.1016/b978-0-443-15274-0.50096-2
- Yayınevi: Elsevier B.V.
- Sayfa Sayıları: ss.607-612
- Anahtar Kelimeler: dynamic PCA, Fault detection, ICA, Manifold Learning, Tennessee Eastman
- Boğaziçi Üniversitesi Adresli: Evet
Özet
In Multivariate Statistical Process Monitoring (MSPM), a large number of measured variables is monitored online usually in a latent variable space. To this end, various linear/ nonlinear, Gaussian/non-Gaussian, and static/dynamic methods have been proposed. While implementations of deep learning methods in MSPM are frequently seen nowadays, less emphasis is given to unsupervised nonlinear manifold learning (ML) methods, such as Isomap. It is important to note that ML takes the geometry of the data into consideration while reducing dimensions, hence may have an advantage, particularly, over Kernelized methods. However, requirement of approximations for out-of-sample (test) points renders ML methods less practical. To remedy this issue, we have recently proposed independent component analysis of polynomial approximation to Isomap embeddings coupled with principal component analysis (ICApIso-PCA) method, and showed that fault detection and isolation performances are drastically improved compared to traditional methods. In the current study, we include lagged process measurements in ICApIso-PCA, i.e. dynamic ICApIso-PCA (dICApIso-PCA), and show that fault detection rate is further increased, and more robust with respect to number of selected components compared to various methods on Tennessee Eastman plant.