Linus Nilsson
	
An introduction to data mining methods CART and MARS with applications
in industrial monitoring

ABSTRACT: We introduce CART---Classification And Regression Trees, and
MARS---Multivariate Adaptive Regression Splines, together with a
methodology for model-based fault detection in industrial plants using
these methods. Factors complicating the analysis of data from
industrial control systems include the presence of non-linear, dynamic
relationships, high dimensionality, large numbers of observations,
limited a priori knowledge, and high requirements regarding
interpretability. CART and MARS can naturally handle many of these
problems. MARS models can be seen as generalized CART models and by
their generalized nature they have continuity properties which are
more in line with the underlying physics. CART is on the other hand
the more scalable method, especially when high order interactions are
present. The two methods also differ regarding how easy it is to make
use of extra a priori knowledge.  These differences are investigated
by applying both methods to real data from a thermal plant. Remarkably
precise prediction of many of the controlled variables in the complex
industrial plant turns out to be possible, using quite limited
physical knowledge.  The results suggest that MARS would be the better
technique of the two for the given purpose.