Automated rule extraction for engine health monitoring

Abstract

A problem of current interest is the automatic classification of potential critical component failures in turbo jet engines. Current processing uses relatively simple metrics or features to measure and characterize changes in sensor data. An alternative solution is to use neural networks coupled with appropriate feature extractors to analyze and automatically extract rules for expert system classifier development. Unfortunately the workings of many neural nets are incomprehensible to humans and thus may be of little utility and not accepted. Elliptical basis function (EBF) neural nets perform classification of input features by clustering and characterizing the feature data with a set of multidimensional basis functions. We have developed a class-dependent EBF neural net to solve this problem. The network is essentially a nearest-neighbor classifier. The network can perform automated rule extraction by examination of the basis functions. Unfortunately, as the number of inputs and the complexity of the neural net grows, the rules generated may become incomprehensible as well. We have used evolutionary programming to select the input feature subset and neural net architecture. The tradeoff is statistical performance versus rule comprehensibility. Here the algorithm is presented as well as results of application to real turbo jet engine data.

This work was supported on a Phase I Small Business Innovative Research contract with the U.S. Air Force / Wright Labs.