ABSTRACT
Andrei V. Gribok, Aleksey M. Urmanov, J. Wesley Hines, and Robert E. Uhrig University of Tennessee, Knoxville, USA
CONTENTS 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 11.2 Collinear, Ill-Posed Problems, Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 11.3 Kernel Regression and MSET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222 11.4 Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 11.5 Data Description and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 11.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Several techniques have been proposed recently for sensor validation in nuclear as well as fossil power plants. They are all based on the same idea of using redundant information contained in collinear data sets to provide an estimation of monitored sensor value. Being data driven statistical techniques they are all prone to the instabilities and inconsistencies caused by collinear finite data sets. This paper examines these techniques from a unifying regularization point of view and presents some experimental comparison of their performance on real plant data. The results show that without proper regularization all the statistical techniques are sensitive to minor variations in the data. Regularization may effectively stabilize the inference making results repeatable and consistent.
