Early Fault Detection in Industrial Plants

Abstract

Modern industrial plants contain enormous numbers of sensors which, in turn, generate enormous amounts of process and diagnostic variable measurements. All this generated data is stored in a Data Historian database and then left untouched. This report evaluates whether there is useful information amongst this unused data, and if so, how this information can best be used to increase the reaction time of plant operators. This is done by examining the application of regression methods to make early faults detection possible. The simulations are performed using historic process data from a crude distiller unit at the Shell Pernis Refinery. Datasets representing both normal and faulty operations are taken from two different subsystems of the crude distiller unit. The output datasets have irregular sampling times that are larger than the input variable datasets so this potential problem is solved by using a linear interpolation to estimate the missing values in the output datasets. The processes in the subsystems are modelled using finite impulse response (FIR) models. Five different regression methods are used to identify these models. This report concludes firstly that the ordinary least squares and ridge regression methods can be used to construct accurate out-of-sample prediction models of key process variables; and secondly, that this can be done without prior process knowledge or extensive process specific analysis.