Automated rule extraction for engine vibration analysis

A problem in engine health monitoring is the automatic detection and classification of potential component failures. Current processing uses simple features to measure and characterize changes in sensor data. An alternative solution uses neural networks coupled with appropriate feature extractors. Unfortunately most neural nets give little insight into the "why" of their output decisions. We have developed a variation of the radial basis function neural net for the problem. The neural net is essentially a nearest neighbor classifier. Classification rules can be found by examination of the basis functions. Rule complexity is reduced by using evolutionary programming to select the input features and neural net architecture. The technique is applied to complex vibration spectral data to yield a simple rule that gives superior performance when compared to a traditional approach. The approach is a valuable tool for developing simple rules when a large feature set is available.