Comparing PCA-based fault detection methods for dynamic processes with correlated and Non-Gaussian variables

Maintenance strategies have been playing an increasingly important role in improving engineering systems' performance, supporting the growth of availability and reliability, and delivering significant savings to their operation. In this scenario, the ability to predict the needs for assets' maintenance at a given future time is one of the major challenges. Consequently, detecting a fault in its early stages becomes a critical part of the development and implementation of an effective maintenance program. In this growing research field, data-based multivariate statistical methods have been standing out over other approaches, with Principal Component Analysis being one of the most-cited methods in the literature. Adaptive and non-adaptive variations of this method have been developed and applied to overcome issues regarding stationarity and data autocorrelation. Accordingly, this article proposes a study in which three adaptive methods combined with three process monitoring metrics are compared to perform early fault detection. Among these methods, two are novelties that incorporate the ability to deal with nonstationary and autocorrelated data. Different types of faults and data distributions are considered and a multi-criteria decision analysis method is applied to define the best alternative among the considered combinations of methods and metrics. The results demonstrate consistency and how the different types of faults can influence the analysis.

Automated summary

Maintenance strategies play a crucial role in improving engineering systems' performance and saving operational costs, with predicting asset maintenance needs and early fault detection being major challenges. Data-based multivariate statistical methods, particularly Principal Component Analysis, stand out for early fault detection. Decision analysis methods can assist in determining the best approach in different scenarios.