ENBIS-16 in Sheffield

11 – 15 September 2016; Sheffield Abstract submission: 20 March – 4 July 2016

Markovian and Non-Markovian Sensitivity Enhancing Transformations for Process Monitoring

12 September 2016, 10:40 – 11:00

Abstract

Submitted by
Tiago Rato
Authors
Tiago J. Rato (University of Coimbra), Marco P. Seabra dos Reis (University of Coimbra)
Abstract
Industrial processes typically present a high degree of correlation among variables, as a result of the natural interactions, control policies and redundant sensors existing in the processing units. This redundant information is useful for tracking process variability and for detecting abnormalities by construction of adequate reference models, being principal component analysis (PCA) a popular methodology in this context. However, PCA only defines the region of acceptable values and does not take into account the process’ causal structure, which is fundamental for fault diagnosis. Other modelling alternatives based on partial correlations can extract the inner causal network underlying data and such information can be further used to model the correlation, namely via sensitivity enhancing transformation (SET). In the simplest form, the SET models each variable by regression on its causal parents, following a Markovian approach. Alternatively, Non-Markovian modelling is here considered by making use of all ancestors and thus potentially improve the description power and robustness of the transform. The transformed variables are then very suitable for detecting mean changes as they highlight deviations from expectation, while reducing the smearing effect problem of PCA. Furthermore, as they are uncorrelated, but still directly related with the original variables, fault diagnosis significantly benefits from such a transform. This is achieved through a procedure similar to the Mason, Tracy and Young (MTY) decomposition, but without the cumbersome computation of multiple contribution terms.
Both SET modelling approaches were compared against PCA methodologies and proved to significantly improve monitoring performance in several cases study.
Acknowledgements:
The authors acknowledge financial support through Portuguese FCT project PTDC/QEQ-EPS/1323/2014 / Compete Project nb. 016658, co-financed by the Portuguese FCT and European Union’s FEDER through the program “COMPETE 2020”.

Return to programme