ICISE Palermo 2016 presentation Tsiamyrtzis .pdf

ICISE-023: A Bayesian Approach for
Online Monitoring of Phase I Data
Konstantinos Bourazas1
1 Dept.

Dimitris Kiagias2

Panagiotis Tsiamyrtzis1

of Statistics, Athens University of Economics and Business, Greece
kbourazas@aueb.gr, pt@aueb.gr

2 School

of Mathematics and Statistics, University of Sheffield, UK
kiagias.dim@gmail.com

Palermo, 20 June 2016

Bourazas, Kiagias, Tsiamyrtzis (AUEB-US)

ICISE-023

Palermo, 20 June 2016

1 / 17

Introduction
In Statistical Process Control/Monitoring (SPC/M) our goal is to
identify as soon as possible when a process moves from the In Control
(IC) to the Out of Control (OOC) state, while we keep the false alarm
rate at a very low (predetermined) level.

Bourazas, Kiagias, Tsiamyrtzis (AUEB-US)

ICISE-023

Palermo, 20 June 2016

2 / 17

Introduction
In Statistical Process Control/Monitoring (SPC/M) our goal is to
identify as soon as possible when a process moves from the In Control
(IC) to the Out of Control (OOC) state, while we keep the false alarm
rate at a very low (predetermined) level.
In frequentist based SPC/M the parameter(s) of interest θ is
considered to be an unknown constant and typically the goal is to
identify transient or persistent shifts of the unknown parameter(s).

Bourazas, Kiagias, Tsiamyrtzis (AUEB-US)

ICISE-023

Palermo, 20 June 2016

2 / 17

Introduction
In Statistical Process Control/Monitoring (SPC/M) our goal is to
identify as soon as possible when a process moves from the In Control
(IC) to the Out of Control (OOC) state, while we keep the false alarm
rate at a very low (predetermined) level.
In frequentist based SPC/M the parameter(s) of interest θ is
considered to be an unknown constant and typically the goal is to
identify transient or persistent shifts of the unknown parameter(s).
Parametric SPC/M control chart methods, like Shewhart control
charts, CUSUM and EWMA, will require the knowledge of the in
control distribution process parameter(s). In practice, this is handled
with the employment of an offline calibration (phase I) period, prior
to the online control/monitoring of the process (phase II).

Bourazas, Kiagias, Tsiamyrtzis (AUEB-US)

ICISE-023

Palermo, 20 June 2016

2 / 17

Introduction

In this talk we will focus in how we can perform online monitoring
during phase I, with emphasis in identifying outliers and propose a
self-starting control/monitoring scheme which will be free from the
phase I requirement.

Bourazas, Kiagias, Tsiamyrtzis (AUEB-US)

ICISE-023

Palermo, 20 June 2016

3 / 17

Introduction

In this talk we will focus in how we can perform online monitoring
during phase I, with emphasis in identifying outliers and propose a
self-starting control/monitoring scheme which will be free from the
phase I requirement.
Phase I (and short run) analysis in frequentist based SPC/M assumes
that we have iid data from the In Control distribution. These data are
used (retrospectively) to estimate the unknown parameter(s) while
testing is performed offline. In case of alarms, the standard iterative
approach is to remove the alarms and recalculate the control limits,
until we get no alarms.

Bourazas, Kiagias, Tsiamyrtzis (AUEB-US)

ICISE-023

Palermo, 20 June 2016

3 / 17

Introduction

In this talk we will focus in how we can perform online monitoring
during phase I, with emphasis in identifying outliers and propose a
self-starting control/monitoring scheme which will be free from the
phase I requirement.
Phase I (and short run) analysis in frequentist based SPC/M assumes
that we have iid data from the In Control distribution. These data are
used (retrospectively) to estimate the unknown parameter(s) while
testing is performed offline. In case of alarms, the standard iterative
approach is to remove the alarms and recalculate the control limits,
until we get no alarms.
The above procedure is known to have certain deficiencies:

Bourazas, Kiagias, Tsiamyrtzis (AUEB-US)

ICISE-023

Palermo, 20 June 2016

3 / 17

Issues in frequentist phase I analysis
Phase I assumes iid data from the in control distribution. What if the
parameter shifts during phase I?

Bourazas, Kiagias, Tsiamyrtzis (AUEB-US)

ICISE-023

Palermo, 20 June 2016

4 / 17

Issues in frequentist phase I analysis
Phase I assumes iid data from the in control distribution. What if the
parameter shifts during phase I?
Phase I needs to be long enough to provide reliable estimates. What
if we have short runs?

Bourazas, Kiagias, Tsiamyrtzis (AUEB-US)

ICISE-023

Palermo, 20 June 2016

4 / 17