Improving the performance of multi-layer perceptrons where limited training data are available for some classes

Abstract

The standard multi-layer perceptron (MLP) training algorithm implicitly assumes that equal numbers of examples are available to train each of the network classes. However, in many condition monitoring and fault diagnosis (CMFD) systems, data representing fault conditions can only be obtained with great difficulty: as a result, training classes may vary greatly in size, and the overall performance of an MLP classifier may be comparatively poor. We describe two techniques which can help ameliorate the impact of unequal training set sizes. We demonstrate the effectiveness of these techniques using simulated fault data representative of that found in a broad class of CMFD problems.