Monitoring multivariate process using improved independent component analysis-generalized likelihood ratio strategy

Though the automation in modern process plants has reduced overall functioning cost, monitoring process for detecting abnormalities is still a challenging task. In this direction, data driven fault detection (FD) methods are commonly applied as they rely purely on historical data collected from a process. Since most practical data are non-gaussian in nature, Independent Component Analysis (ICA) strategy has found to be the right way for non-gaussian process monitoring. To determine faults, generalized likelihood ratio test (GLRT) has been merged with other data driven approaches. In this paper, we propose a FD strategy where ICA is used as modeling framework and generalized likelihood ratio test (GLR) as a fault detection index. Once ICA model is established, uncorrelated residuals obtained from the model would be evaluated by the GLR test to detect faults. The monitoring performance is compared with traditional ICA based fault indicators- I²_d, I²_e and SPE statistics. In addition to fault detection, fault isolation strategy is also presented in this work. The proposed ICA based GLR strategy performance is evaluated using a simulated quadruple tank process and benchmark Tennessee Eastman process. The simulation results shows that the proposed fault detection approach is able to detect sensor faults very effectively.