[Research] there WILL be a lab meeting tomorrow!

[Kristen Schrauder]

Just a reminder, there will be a lab meeting today (2/23) in NSH 1507 at
4:30pm.  Thanks!

-----Original Message-----
From: Kristen Schrauder [mailto:kristens at cs.cmu.edu] 
Sent: Wednesday, February 22, 2006 1:37 PM
To: 'research at autonlab.org'
Cc: 'awd at cs.cmu.edu'
Subject: RE: there WILL be a lab meeting tomorrow!

We will be in NSH 1507 as usual and there will be food :)

-----Original Message-----
From: Artur Dubrawski [mailto:awd at cs.cmu.edu] 
Sent: Wednesday, February 22, 2006 10:44 AM
To: research at autonlab.org
Cc: Kristen Schrauder
Subject: there WILL be a lab meeting tomorrow!

Hello,

I am sorry for the late notice,
but in fact we will have a lab
meeting this Thursday and a guest 
speaker who, without much arm-twisting, 
agreed to talk about the topics explained
below.

Kristen: please make sure we have room
available and let everyone know if it
won't be the usual NSH 1507. And please
get us some food :-)

Artur

-----
Piero Bonissone, General Electric Global Research

"Knowledge, Time and Decisions (How, When, and What): 
A Framework for Computational Intelligence Applications in Predictive
Health Management"

This effort presents a framework for understanding the requirements,
limitations, and performance of 
a large variety of Computational Intelligence (or Soft Computing) models
- based on Fuzzy Sets 
(Instance-based, C-means) evolutionary (GA, MOEA, GP), Neural Nets
(Feedforward NN, ANFIS,SOM), 
Statistical (Random Forests, Kernel-Based approaches), Information
Theory (Kolmogorov Complexity), etc.

We apply these models to the area of Prognostics and Health Management
(for equipment such as aircraft engines, 
CT scanners, locomotives, etc.).  To analyze the output (decisions) of
the models, we focus on two factors:
1) the time horizon for the decision (tactical, operational, strategic)
and
2) the degree of availability of domain knowledge to construct the
model.

For the latter we use linguistics as a metaphor.
We analyze the progression from simple lexicon to annotated lexicon,
morphology, syntax, semantics, 
and pragmatics and compare it with the injection of domain knowledge in
PHM. 
This knowledge start with event messages and is extended with event
messages taxonomy, 
normal/failure labels, failure signatures, key variables (features) and
first principle relationships, 
and context-based model selection. 

>From the combination of data and knowledge we can 
perform anomaly detection, anomaly identification, failure mode analysis
(diagnostics), estimation of remaining useful life (prognostics), and
optimal health management decisions.
We illustrate key combinations of this space (Time x Knowledge) with
short problem descriptions and related SC-based solutions.