Autoencoder Based Residual Generation for Fault Detection of Quadruple Tank System

Abstract

Increasing complexity of industrial processes has made statistical methods for process monitoring and diagnosis a more attractive alternative to model-based methods. A primary reason is that statistical approaches can be formulated to rely less on process knowledge. Since multivariable processes can exhibit complex, nonlinear dynamics, there is a need for methods capable of diagnosing nonlinear process data. A Monte Carlo simulation was conducted on a numerical model of the quadruple tank process (QTP) - a novel multivariate nonlinear process. The simulation was designed so that the QTP exhibited bipartite nonlinear behavior. Reference data obtained from the simulation was used to obtain principal component analysis (PCA) and autoencoder (AE) models. The models generated residuals that were used to monitor the condition of the process. The results showed that AEs, which have nonlinear functionalities, performed better than PCA models at generating residuals.