Wear Debris Identification with Neural Networks

A feedforward neural network is applied to identification of wear debris generated under different sliding conditions. In order to describe characteristics of debris of various shapes and sizes, four representative parameters for groups of randomly sampled wear debris, i. e. 50% volumetric diameter, average elongation, average roundness and average reflectivity, are used as inputs to the network. Debris sampled in five steel sliding experiments are chosen as examples. It is shown that identification results depend on the ranges of these parameters, and that the ranges are determined by the sample size used for averaging. In the present case, a sample size of fifty pieces of debris provides a satisfactory identification result with less than ten percent error. It is also demonstrated that use of data for larger particles leads to better results. We discuss how the network determines difference in debris features, and how this approach can be applied to diagnosis of sliding surfaces.