We propose a software package that combines recent advances in several pattern recognition techniques, including neural networks, multivariate statistics and model-based reasoning. Results from these prognostics will be combined into a single probability of failure for the modeled component or subsystem of the ABL, which would be output to a simple graphical user interface (GUI). The preliminary software design suggested contains modules that are highly cohesive and decoupled, particularly with respect to the prognostic algorithms and the library specific to the ABL system. This decoupling greatly enhances the reuse of the software for other components or subsystems of the ABL, or for domains completely separate from the ABL. The work that we propose seeks not only to establish the feasibility of the approach that we will take, but will also create a solid software design for future work. In addition, at least one prognostic algorithm will be completed along with a rudimentary library relating to a single ABL component or subsystem. The completed algorithm(s) along with the library will be combined in a software package that will generate failure predictions on the modeled component or subsystem. Anticipated Benefits/Commercial Applications: There is a great need for prognostic systems throughout the MDA, DoD, and commercial markets. As systems grow increasingly complex, human analysis of these systems becomes less feasible. The first commercialization attempts will be with MDA and the ABL system. Phase one and Phase two funding only allows for analysis of a very limited part of the ABL system, and the partnership of BKA and APL will continue to pursue additional contracts to continually improve the prognostic systems, and the expand the domain that is being monitored. In addition to the ABL system, the software can be marketed throughout the MDA, as the decoupling of the domain libraries from the algorithms and program operation will allow rapid and inexpensive development for other MDA systems. Once completed, the software will allow prognostics on nearly any system with real time data flows, only requiring the development of new domain libraries. Outside of the MDA, the DoD community at large has tremendous need for automated prognostics of mechanical systems, and these markets will be sought as well. In additional to mechanical systems, however, software and information systems can also be diagnosed with such a tool. BKA will continue to require and develop prognostic systems for its tactical software effort, and this tool will be integrated into such general-purpose software as JEBRA, as well as specific software packages for more specialized domains. The software package itself as well as the lessons learned from the feasibility study will be invaluable tools as BKA continues these efforts. APL will use this tool in a variety of software efforts for the DoD and NASA. The aforementioned OAK software package, for instance, could be evolved to include prognostics to the already functioning diagnostic, automated response, and control system to reduce manning on Navy vessels. Additionally, this software would likely be integrated into sensor grids to enhance the survivability and reliability of those grids. Finally, APL would seek to use this software package to add prognostics to unmanned underwater and aerial vehicles to predict failures in these systems. The need for automated prognostics is not unique to the DoD, however. A pervasive problem in the commercial sector is to optimize manufacturing processes and detect and prevent failures. The immense cost of individualized software development prevents many such companies from pursuing such solutions however. Using adaptations of this software, BKA will seek to provide solutions to such problems to consumers in the commercial market at a substantially reduced rate.
Keywords: neural networks, model-based reasoning, aritificial intelligence, risk assessment, multivariate statistics, data-driven prognostics, mathematical reasoning, system reliability
