Classification of wear debris using a neural network

Abstract

In this work, an evaluation was made to prove the possibility of employing a neural net method for the classification of debris and monitoring of a lubricated contact pair. We trained a neural net to classify the severity of wear into two types from the morphological features of the wear debris. The following procedures were used. First, the shape of wear particles was characterized by Fourier descriptors. The Fourier descriptors were considered as coordinates of a point in multidimensional feature space. A set of points form a cluster, and the location and structure of the cluster depend on the morphology of the wear particles and the current conditions of the contact system. A distance distribution between the debris in the feature space was used to represent the location of the cluster. Second, we trained a back-propagation neural net. To train the neural net, we used the distance distribution corresponding to the different stages of the wear process as an input vector and encoded the wear rate as a desired response. The network was then further trained until the desired error goal was achieved. Finally, we tested the trained neural net. The ability of the neural net method to monitor wear is shown. (C) 1997 Elsevier Science S.A.