Advance Failure Warning via Data Driven Stochastic Models

The ideal prognostic system would be generic, requiring no domain-specific knowledge for its application. The prognostic system must also perform across all operating regimes without generating an excessive number of false alarms, and must successfullydiagnose problems. The semi-empirical technique proposed herein by Barron Associates, Inc. and its academic partners at Princeton University takes advantage of sophisticated black-box modeling, efficient nonparametric state estimation, and sensitivestatistical change detection and isolation algorithms. Observers composed of banks of unscented particle filters will be used to estimate the internal states of the system that is being monitored, after a model has been constructed from preliminary data.These resulting state estimates will then be used to synthesize output signals that are insensitive to noise and unmodelled dynamics, but sensitive to faults. A statistical change detection technique (based on a modification of the standard generalizedlikelihood ratio statistic) is then used to process these output signals to detect faults in real time. Detection of anomalous behavior is made subject to user-specified probabilities for false alarms and missed detections. Analytical redundancy andprobabilistic modelling will be used to ensure the correct diagnosis and isolation of problems. The ability to predict machine/equipment events has significant commercial potential in aircraft, power, manufacturing, processing, transportation, and otherindustrial applications where such capability would allow companies to improve reliability and safety, reduce downtime, and lower the direct maintenance cost of physical assets.