Improved fault detection using Dynamic Independent Component Analysis (DICA): An application to multi-variate system

Abstract

In this paper, a statistical multi-variate technique based on Dynamic Independent Component Analysis (DICA) is proposed for monitoring abnormalities in a chemical process. Multi-variate fault detection (FD) technique based on Principal Component Analysis (PCA) is restricted in capturing gaussian features of industrial data and it also assumes that observations at present time instant are not dependent on previous time instant. These assumptions do not apply for industrial processes due to the random characteristics of the variables and the underlying dynamics of the process. Another multi-variate FD technique named Indendent Component Analysis (ICA) has the ability of representing data as a function of latent variables (IC‘s) which are independent and this assumption is crucial to capture non- gaussian features in the data. The dynamics in the process data could be incorporated through dynamic ICA modeling where ICA model is embedded with lagged variables for capturing plant dynamics. In the current work, dynamic ICA (DICA) is used as the modeling frame-work while I2d, I2e and SPE statistics are the fault detection indicators. In ICA model development, the conventional FastICA algorithm involves random initialization of matrix B which results in different solutions for each iteration. To avoid this concern, in the current work, the matrix B is be initialized to a identity matrix to provide constant solution in each iteration. The performance of developed DICA strategy is demonstrated on a multi-variate process and a simulated quadruple tank process. The simulation results clearly suggest that the DICA strategy is able to detect anomalies effectively.