Improved Fault Detection and Isolation Using Enhanced Multiscale Principal Component Analysis: Algorithms and Applications
Abstract
Effective and reliable fault detection and isolation (FDI) methods are essential for the efficient operation of industrial processes. This work has developed a data-driven algorithm called Enhanced Multi-Scale PCA (EMSPCA) to improve fault detection performance of the conventional MSPCA method. It also extends EMSPCA to isolation by utilizing a PCA reconstruction-based approach to improve the fault isolation performance.
A critical analysis presented in this work, shows that the conventional MSPCA detection rate is obstructed by its inaccurate predictions of detection thresholds. To address this issue, EMSPCA alters the way wavelet coefficients are processed in the training and testing data, such that, the predicted threshold is suitably tighter for smaller fault projections, which ensures a much better detection rate. A soft-thresholding technique is also implemented to ensure the false alarm rates remain low. Moreover, this research extends the EMSPCA method to account for isolation at multiple scales. Previous research has used contribution plot isolation approaches in the multiscale framework, but here, the reconstruction-based approach is employed. Reconstruction-based approaches suffer less from the smearing effect and can therefore achieve better isolation rates. Smearing occurs when one variable contaminates or “smears” another variable’s contribution or isolation index to the point of misdiagnosis. The work will investigate how the multiscale PCA framework can minimize the amount of smearing to produce optimal isolation performance. It will also offer a comparison between contribution plot and reconstruction-based isolation performances and present the impact of decimated and undecimated wavelet transforms on detection and isolation performances.
To obtain statistical and meaningful conclusions, a randomized synthetic linear model with an embedded shift-in-the-mean univariate fault is utilized. Monte Carlo simulations are used to evaluate the false alarm, detection rates, and isolation rates across all decomposition depths and a range of fault sizes. To further validate the algorithm and the FDI improvements it realizes, this work will utilize real data from a pilot distillation plant and two TEP units with a fabricated sensor fault embedded. These results will demonstrate the superior FDI performance of the EMSPCA reconstruction-based approach.
Subject
fault isolationPCA
MSPCA
fault detection
FDI
data
Process Monitoring
Multiscale
wavelet analysis
Citation
Malluhi, Byanne Qutaibah (2019). Improved Fault Detection and Isolation Using Enhanced Multiscale Principal Component Analysis: Algorithms and Applications. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /189104.