Dataset Identification:
Resource Abstract:
- An optimal alarm system is simply an optimal level-crossing predictor that can be designed to elicit the fewest false alarms
for a fixed detection probability. It currently use Kalman filtering for dynamic systems to provide a layer of predictive
capability for the forecasting of adverse events. Predicted Kalman filter future process values and a fixed critical threshold
can be used to construct a candidate level-crossing event over a predetermined prediction window. Due to the fact that the
alarm regions for an optimal level-crossing predictor cannot be expressed in closed form, one of our aims has been to investigate
approximations for the design of an optimal alarm system. Approximations to this sort of alarm region are required for the
most computationally efficient generation of a ROC curve or other similar alarm system design metrics. Algorithms based upon
the optimal alarm system concept also require models that appeal to a variety of data mining and machine learning techniques.
As such, we have investigated a serial architecture which was used to preprocess a full feature space by using SVR (Support
Vector Regression), implicitly reducing it to a univariate signal while retaining salient dynamic characteristics (see AIAA
attachment below). This step was required due to current technical constraints, and is performed by using the residual generated
by SVR (or potentially any regression algorithm) that has properties which are favorable for use as training data to learn
the parameters of a linear dynamical system. Future development will lift these restrictions so as to allow for exposure to
a broader class of models such as a switched multi-input/output linear dynamical system in isolation based upon heterogeneous
(both discrete and continuous) data, obviating the need for the use of a preprocessing regression algorithm in serial. However,
the use of a preprocessing multi-input/output nonlinear regression algorithm in serial with a multi-input/output linear dynamical
system will allow for the characterization of underlying static nonlinearities to be investigated as well. We will even investigate
the use of non-parametric methods such as Gaussian process regression and particle filtering in isolation to lift the linear
and Gaussian assumptions which may be invalid for many applications. Future work will also involve improvement of approximations
inherent in use of the optimal alarm system of optimal level-crossing predictor. We will also perform more rigorous testing
and validation of the alarm systems discussed by using standard machine learning techniques and consider more complex, yet
practically meaningful critical level-crossing events. Finally, a more detailed investigation of model fidelity with respect
to available data and metrics has been conducted (see attachment below). As such, future work on modeling will involve the
investigation of necessary improvements in initialization techniques and data transformations for a more feasible fit to the
assumed model structure. Additionally, we will explore the integration of physics-based and data-driven methods in a Bayesian
context, by using a more informative prior.
Citation
- Title Optimal Alarm Systems
-
- revision Date
2014-12-12T13:44:20
Resource language:
[u'en-US']
Constraints on resource usage:
-
- Constraints
-
- Use limitation statement:
- public
point of contact
-
publisher
- individual Name {u'hasEmail': u'mailto:rodney.martin@nasa.gov', u'fn': u'RODNEY MARTIN'}
- organisation Name
{u'subOrganizationOf': {u'subOrganizationOf': {u'name': u'U.S. Government'}, u'name': u'National Aeronautics and Space Administration'},
u'name': u'Dashlink'}
-
- Contact information
-
-
- Address
-
- electronic Mail Address
Back to top:
Metadata data stamp:
2014-12-12T13:44:20
Metadata contact
-
publisher
Metadata scope code
dataset
Metadata standard for this record:
ISO 19115:2003 - Geographic information - Metadata
standard version:
ISO 19115:2003
Metadata record identifier:
DASHLINK_119
Metadata record format is ISO19139 XML (MD_Metadata)