ENBIS-16 in Sheffield

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

Extreme Value Analysis for Multivariate Process Monitoring

14 September 2016, 09:00 – 09:20

Abstract

Submitted by
Stijn Luca
Authors
Stijn E. Luca (KU Leuven), David A. Clifton (University of Oxford), Bart Vanrumste (KU Leuven)
Abstract
Process monitoring is a diverse field that is applicable to the monitoring of machines, industrial production processes and health monitoring. Fault detection is a crucial part of process monitoring, where one wishes to identify faults or abnormalities that indicate a deviation from the normal state of the system.
We will treat a problem related to the detection of faults in multivariate processes. Most of the literature on multivariate process monitoring deals with point-wise approaches evaluating individual points of the process. This work treats the more general problem of modelling patterns of measurements from a process. In particular we consider the general problem of determining whether a given set of d-dimensional data points S={x_1,x_2,…,x_n} is coming from a statistical distribution X modelling the “normal” state of a process.
To this end, it is shown that a models based on extreme value statistics can be used for multivariate process monitoring, that enables to fuse different types of information hidden in low-density regions of data space. This is particular beneficial when data from the faulty situation is sparse such that a model of the abnormal state is absent and other classification techniques are sub-optimal.
The method is illustrated on a healthcare application where epileptic patients are monitored during their daily routine. Treating epileptic seizures as extreme or statistically rare allow to detect them in a set of acceleration data acquired from a wristband.
View paper

Return to programme