Collective Mind: Continuous, Automatic Learning to Improve ...

unknownlippsΤεχνίτη Νοημοσύνη και Ρομποτική

16 Οκτ 2013 (πριν από 4 χρόνια και 2 μήνες)

77 εμφανίσεις

Collective Mind:


Continuous, Automatic Learning to
Improve Equipment Maintenance

A Request for Guidance

CBM+

February 11, 2005


Norman Sondheimer Al Wallace Peter Will


UMass Amherst RPI USC/ISI






































Vision: New Paradigm for

Maintenance Decision Support

Objective: Actively Manage the Maintenance Process


Continuously Improve Planning, Response, and Execution


Operating at All Levels, All Phases of Operations


Linking all Elements into a Living, Distributed, Global Maintenance System

Basic Building Block:

Self
-
Aware Platform

Global Community of Continuously Improving

Equipment

Sense and Respond

Maintenance Network

Technical Approach...

Collective Mind:

Communities of

Self
-
Aware Platforms

Self
-
Sustainment Requirement


Army Future Force units in 2020


Air Expeditionary Force


Navy/Marines Sea Basing

Critical Capabilities: Reliability

Prognostics

Army: “The UA is self
-
sustainable for 3
-
7 days of
operations and maintains combat power with
dramatically reduced theater stockpiles.”


Challenges for Prognostics


Lack of Physics of Failure models


Missing sensor suites


Low equipment
-
utilization rates


etc.


But Maintenance Crews


Have the ability to improve reliability of their
equipment over time


We have yet to tap the data we have!

Claim: Existing field experience can
be used to improve Prognostics


Discover similar units


Peers form a “Collective”


Evaluate unit under consideration using
experience of the Collective


Improve discovery and evaluation
methods based on weapon systems
success


Learning gives us our title “Collective Mind”

Key Technology: Statistical Machine Learning

Example: Locomotive Selection


The Mission:


Select 12 Locomotives


to go from CA to PA

Design and Configuration


Type


Electrical System




Utilization Information


Age


Mileage



Average miles/day





Maintenance Information


Time elapsed since last repair


Median time between repairs


Median time from repair to next


recommendation (Rx)




Decision Support Data:


200+ Basic Parameters

Data from GE Transportation

Locomotives Network Enabled

Identifying Peers

f2: miles/day

f1: Rx/Year

Peers of Unit 5700

Unit 5700

f3: ??

P1

P2

P3

P4

P6

P5

P7

= Peers of Unit 5700

Collective:
Peers with similarity measure

Peer experience forms Mission Reliability (MR) rank

Learning:
similarity measure updated by accuracy of MR Rank


State of the Practice:




non
-
Machine Learning

Selection Criteria

% of Correctly Classified
Units: Top 20%

(Sample Performance)

Lowest Mileage

17%

Newest Units

18%

Random

20%

Highest Energy (MWHRS) generated

24%

Highest Miles/ Hours Moving

26%

Highest Percentage Hours Moving

29%

Lowest Percentage of Failures in Most
Critical Subsystem

38%

Lowest Ratio: Recommendations / Age

49%

Accuracy and Robustness of the Peer
Approach

Excellent Performance with
Existing Sensors on Legacy
Systems

48.1%

55.8%

63.5%

32%

37%

37%

20%

8%

6%

5.0%

15.0%

25.0%

35.0%

45.0%

55.0%

65.0%

75.0%

1

2

3

Time Slices

Selection Performance

Learned Peers

non
-
Machine Learning

Random

(52 out of 262 units)

(52 out of 634 units)

(52 out of 845 units)

50%

46%

44%

39%

37%

35%

35%

32%

30%

54%

50%

48%

48%

44%

47%

44%

41%

44%

43%

45%

0.25

0.3

0.35

0.4

0.45

0.5

0.55

Performance on Selection Task

Non
-
ML

Peers

Highest contributing parameters
assumed missing

Learned Peers show better performance &
continuous improvement

Robust to Missing
Information

Guidance


Possible Futures


Test in of Current Technology in Services


Proposals with AFRL to eLog21: F110 or F100


Partnership with AMCOM, Fort Rucker:
Blackhawk


Development of Technology


Meeting with ONR and Marines ALP


Development of Vision


Meeting with DARPA IXO

Questions?


Collective Mind


Continuously Improving



Equipment Maintenance

Norman Sondheimer

U of Massachusetts Amherst

Amherst, MA 06001

sondheimer@som.umass.edu

Al Wallace

RPI

Troy, NY 12180

wallaw@rpi.edu

Peter Will

USC/ISI

Marina del Rey, CA

will@isi.edu

Piero Bonissone,

Steve Linthicum, Anil Varma

GE Global Research

Schenectady, NY 12301

linthicu@crd.ge.com