Selecting Data from Plant Database

4.3 Conclusions

From authors experience, confirmed also from applications not reported here, it can be concluded that the task of outlier detection is a process that cannot be totally automated and all available information, including expert knowledge, has to be taken into account. Moreover, a fundamental role in the choice of outlier detection criterion is played by the final objective of the analysis.

As a general rule if a careful modeling of system dynamics is of interest, methods that exploit only statistical inspection of process data can have an undesired conservative effect. They tend to remove peaks which can carry precious information about system dynamics. A better choice in these cases is methods like PLS that take into account input-output relationships. The results obtained should be in any case carefully investigated due to the rough nature of the outlier detection process. The suggested methods, in fact, limit the search process to the case of linear static correlations, while their nonlinear extensions are usually computationally expensive. If a coarse model is of interest, a rule of thumb is that in the case of an uncertain candidate, it is better to eliminate it, instead of using wrong information in the process of model identification. Of course the number of available data can make the designer more or less sensitive to this rule.