0

Full Content is available to subscribers

Subscribe/Learn More  >

Machine Learning and Model Based Reasoning for Prognostics of Complex Systems

[+] Author Affiliations
Kenneth Marko

ETAS Group

Paper No. IMECE2005-81625, pp. 793-800; 8 pages
doi:10.1115/IMECE2005-81625
From:
  • ASME 2005 International Mechanical Engineering Congress and Exposition
  • Manufacturing Engineering and Materials Handling, Parts A and B
  • Orlando, Florida, USA, November 5 – 11, 2005
  • Conference Sponsors: Manufacturing Engineering Division and Materials Handling Division
  • ISBN: 0-7918-4223-1 | eISBN: 0-7918-3769-6
  • Copyright © 2005 by ASME

abstract

Model based reasoning (MBR) has been shown to be an effective means of providing condition based maintenance for many high-value assets for which accurate first principle models have been developed. Yet, many low-cost complex computer controlled systems are mass-produced without the concurrent provision of precise physics based models. We wish to utilize new developments in machine learning coupled with model based reasoning methods to address this deficiency. In particular, we shall demonstrate that for an important class of these systems, the extremely large number of mass produced, complex engine systems which power vehicles and small power generation plants, effective means of providing MBR for condition based maintenance exists. It will be recognized that the methodology also has much broader applicability. We will show that a class of dynamic neural networks can be used to provide high-fidelity models of these complex systems that permit an analysis of differences between predicted normal behavior and actual plant behavior to be analyzed to detect deviations from nominal behavior which will be shown to be valuable in estimating time-to-failure for such systems. The realization of this capability is dependent upon the development of extremely efficient and powerful training algorithms for these dynamics neural networks. While many simple training schemes have been in use for many years, they generally fail to provide the needed model accuracy when they are applied to training the relatively “large” multi-layered dynamic networks that are needed to precisely mimic plant behavior over all operating conditions. Our approach has several advantages over these simpler, but less effective methods. Three major improvements are the rate at which learning proceeds, the provision of a means to optimize the learning rate through-out the process, and the dramatic improvements observed in learning in the final stages of training when the error feedback from training examples are extremely small and the associated error covariance matrices almost vanish. We shall demonstrate with data drawn from production vehicles, that for several important problems in analyzing system performance in these vehicles, sufficient model fidelity can be attained to meet the requirements on detection efficiency, false alarm immunity and alarm response time which are required for effective diagnostics and prognostics. Finally we shall discuss the manner in which the deviations are analyzed to not only identify that a failure has been detected but also the means by which the probable root cause may be isolated.

Copyright © 2005 by ASME

Figures

Tables

Interactive Graphics

Video

Country-Specific Mortality and Growth Failure in Infancy and Yound Children and Association With Material Stature

Use interactive graphics and maps to view and sort country-specific infant and early dhildhood mortality and growth failure data and their association with maternal

NOTE:
Citing articles are presented as examples only. In non-demo SCM6 implementation, integration with CrossRef’s "Cited By" API will populate this tab (http://www.crossref.org/citedby.html).

Some tools below are only available to our subscribers or users with an online account.

Related Content

Customize your page view by dragging and repositioning the boxes below.

Related eBook Content
Topic Collections

Sorry! You do not have access to this content. For assistance or to subscribe, please contact us:

  • TELEPHONE: 1-800-843-2763 (Toll-free in the USA)
  • EMAIL: asmedigitalcollection@asme.org
Sign In