Characterization of wear particles and their relations with sliding conditions

Abstract

Wear particles and worn surfaces generated in pin-on-disk steel sliding experiments are studied by microscope image analysis and two types of neural networks. Features of wear particles described by four particle descriptors depend strongly on sliding conditions. A multi-layer neural network successfully learns the relations between wear particle features and sliding conditions. If the network is trained with data representing typical features, it also recognizes the particles having similar features. This suggests that the network can be used as a tool for condition monitoring in which the network identifies wear particles produced under unknown sliding conditions to predict that conditions. A self-organizing neural network using the competitive learning rule classifies the wear particles based on their features without any supervisor data. Particle features are expressed by the position on a two-dimensional feature map. This type of network is useful in finding typical particle features, which in turn can be used as supervisor data for the multi-layer neural networks. In another application of the self-organizing network, microscopic images of both wear particles and worn surfaces are automatically classified, and characteristics of each surface are represented by the distribution of weights on the feature map.