Many natural or complex engineered systems rely upon critical functions or processes that can be measured with the aid of various sensors or other novel devices. As a result, sensor and measurement data can be used to learn a parametric or non-parametric model of the behavior for a given process or metric. For such processes or metrics, it may be critical to avoid or be forewarned of impending level crossings that may characterize entry into extreme or potentially catastrophic operating regimes. Under certain circumstances, the metric to be monitored may represent the residual, or difference between an actual value and a predicted value generated by an independent regression method, rather than a physical process having a physically interpretable meaning.

The state of the art in the prediction or forecasting of adverse events with respect to residual-based event detection is currently based upon the use of methods that have been derived from a technique called MSET (Multivariate State Estimation Technique). This technique was originally developed at Argonne National Laboratory, but has since been adapted for myriad applications spanning a broad range of disciplines. Some disadvantages of this method lie in the fact that certain technical conditions are required for the implementation of the detection portion of the method. Furthermore, the regression portion of methods derived from MSET has a limited ability to reduce residual error, and lacks robust numerical stability properties and controls for complexity, which may be provided for with alternate regression techniques, some of which may be more scalable.

ACCEPT consists of an overall software infrastructure framework and two main software components. The software infrastructure framework consists of code written to preprocess data, pass information between the two main software components, learn models that will be shared by nearly all of the elements in one of the two software components (which will require calling third-party open source software modules), and select which element/method should be used in each one of the two main software components. The two main software components can use interchangeable software elements that enable the regression and detection functionality. Software elements can either be distributed with the initial release, or called separately as independent elements that have been open-sourced already.

A novel feature is relaxation of the technical conditions required for robust early detection of the onset of adverse events. Rather than assume the residuals are white, or have been pre-whitened either with the aid of an optimal filter or by the use of an appropriate regression technique, serial correlations are retained to be learned using applicable data-driven or machine learning methods. Furthermore, the method attempts to characterize an adverse event via definition through various testable hypotheses that can be selected by the user, one of which is based upon an extreme value level crossing over a predefined prediction horizon, and the other an abrupt change in model parameters as is performed with MSET.

This work was done by Rodney Martin, Ashok Srivastava, and Nikunj Oza of Ames Research Center; Santanu Das and Vijay Janakiraman of The Regents of the University of California Santa Cruz; Bryan Matthews of Stinger Ghaffarian Technologies Inc.; and Richard Watson of Perot Systems Government Services. NASA is seeking partners to further develop this technology through joint cooperative research and development. For more information about this technology and to explore opportunities, please contact David Morse at This email address is being protected from spambots. You need JavaScript enabled to view it. or 650-604-4724. ARC-16839-1

NASA Tech Briefs Magazine

This article first appeared in the January, 2017 issue of NASA Tech Briefs Magazine.

Read more articles from this issue here.

Read more articles from the archives here.