Gated Experts Neural Networks for Prognostics

The prognostic tasks in component health monitoring require recognizing early signs of system malfunctioning and recommending whether a maintenance procedure should be carried out. A prognostic system needs to separate nominal component behavior from afaulty ones even in the cases that those behaviors are similar. Advanced pattern recognition techniques are required to separate nominal and faulty input-output component data vectors in a complex high-dimensional space. We propose to develop a GatedExperts based Neural Network for Prognostics (GE-NNP) architecture and test its performance on a subsystem of the ABL system. GE-NNP is based on the integration of the analysis of three local experts that are developed using the Extended Auto AssociativeNeural Networks (E-AANN), Kohonen Self-Organizing Maps (KSOM), and Radial Basis Function based Clustering (RBFC). The Gated Experts approach in GE-NNP combines the recommendations of E-ANN, KSOM, and RBFC to dynamically give more weight to the ones bettersuited for the current context. In Phase I we will use data from the airborne laser (ABL) program to evaluate the performance of GE-NNP for detecting anomalies in determining how much a certain input-output data observation deviates from what is expectedfor a healthy component. Will detect potential faults and will warn the users well ahead of time to perform maintenance/repair operations thereby saving millions of dollars for industrial repairs and maintenance.