Data Mining In A Vibration Analysis Domain By Extracting Symbolic Rules From Rbf Neural Networks

Abstract

Neural networks are becoming an increasingly popular technique for modelling data with complex and/or non-linear relationships. Diagnostic systems for condition monitoring applications fall particularly into this category, especially those using spectral vibration data. However, neural networks do have some major disadvantages compared with rule based diagnostic systems. The most important criticism is the lack of any explanation system, which would open up the neural networks internal operation for scrutiny. This paper illustrates how the internal parameters of an RBF network can be converted into symbolic rule format. The rule extraction algorithm and is described in detail